Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
75bb4625 JA |
2 | /* |
3 | * Block multiqueue core code | |
4 | * | |
5 | * Copyright (C) 2013-2014 Jens Axboe | |
6 | * Copyright (C) 2013-2014 Christoph Hellwig | |
7 | */ | |
320ae51f JA |
8 | #include <linux/kernel.h> |
9 | #include <linux/module.h> | |
10 | #include <linux/backing-dev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blkdev.h> | |
fe45e630 | 13 | #include <linux/blk-integrity.h> |
f75782e4 | 14 | #include <linux/kmemleak.h> |
320ae51f JA |
15 | #include <linux/mm.h> |
16 | #include <linux/init.h> | |
17 | #include <linux/slab.h> | |
18 | #include <linux/workqueue.h> | |
19 | #include <linux/smp.h> | |
e41d12f5 | 20 | #include <linux/interrupt.h> |
320ae51f | 21 | #include <linux/llist.h> |
320ae51f JA |
22 | #include <linux/cpu.h> |
23 | #include <linux/cache.h> | |
24 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 25 | #include <linux/sched/topology.h> |
174cd4b1 | 26 | #include <linux/sched/signal.h> |
320ae51f | 27 | #include <linux/delay.h> |
aedcd72f | 28 | #include <linux/crash_dump.h> |
88c7b2b7 | 29 | #include <linux/prefetch.h> |
a892c8d5 | 30 | #include <linux/blk-crypto.h> |
320ae51f JA |
31 | |
32 | #include <trace/events/block.h> | |
33 | ||
34 | #include <linux/blk-mq.h> | |
54d4e6ab | 35 | #include <linux/t10-pi.h> |
320ae51f JA |
36 | #include "blk.h" |
37 | #include "blk-mq.h" | |
9c1051aa | 38 | #include "blk-mq-debugfs.h" |
320ae51f | 39 | #include "blk-mq-tag.h" |
986d413b | 40 | #include "blk-pm.h" |
cf43e6be | 41 | #include "blk-stat.h" |
bd166ef1 | 42 | #include "blk-mq-sched.h" |
c1c80384 | 43 | #include "blk-rq-qos.h" |
320ae51f | 44 | |
f9ab4918 | 45 | static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); |
c3077b5d | 46 | |
34dbad5d OS |
47 | static void blk_mq_poll_stats_start(struct request_queue *q); |
48 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
49 | ||
720b8ccc SB |
50 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
51 | { | |
3d244306 | 52 | int ddir, sectors, bucket; |
720b8ccc | 53 | |
99c749a4 | 54 | ddir = rq_data_dir(rq); |
3d244306 | 55 | sectors = blk_rq_stats_sectors(rq); |
720b8ccc | 56 | |
3d244306 | 57 | bucket = ddir + 2 * ilog2(sectors); |
720b8ccc SB |
58 | |
59 | if (bucket < 0) | |
60 | return -1; | |
61 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
62 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
63 | ||
64 | return bucket; | |
65 | } | |
66 | ||
3e08773c CH |
67 | #define BLK_QC_T_SHIFT 16 |
68 | #define BLK_QC_T_INTERNAL (1U << 31) | |
69 | ||
f70299f0 CH |
70 | static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q, |
71 | blk_qc_t qc) | |
72 | { | |
73 | return q->queue_hw_ctx[(qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT]; | |
74 | } | |
75 | ||
c6699d6f CH |
76 | static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx, |
77 | blk_qc_t qc) | |
78 | { | |
efbabbe1 CH |
79 | unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1); |
80 | ||
81 | if (qc & BLK_QC_T_INTERNAL) | |
82 | return blk_mq_tag_to_rq(hctx->sched_tags, tag); | |
83 | return blk_mq_tag_to_rq(hctx->tags, tag); | |
c6699d6f CH |
84 | } |
85 | ||
3e08773c CH |
86 | static inline blk_qc_t blk_rq_to_qc(struct request *rq) |
87 | { | |
88 | return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) | | |
89 | (rq->tag != -1 ? | |
90 | rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL)); | |
91 | } | |
92 | ||
320ae51f | 93 | /* |
85fae294 YY |
94 | * Check if any of the ctx, dispatch list or elevator |
95 | * have pending work in this hardware queue. | |
320ae51f | 96 | */ |
79f720a7 | 97 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 98 | { |
79f720a7 JA |
99 | return !list_empty_careful(&hctx->dispatch) || |
100 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 101 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
102 | } |
103 | ||
320ae51f JA |
104 | /* |
105 | * Mark this ctx as having pending work in this hardware queue | |
106 | */ | |
107 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
108 | struct blk_mq_ctx *ctx) | |
109 | { | |
f31967f0 JA |
110 | const int bit = ctx->index_hw[hctx->type]; |
111 | ||
112 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
113 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
114 | } |
115 | ||
116 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
117 | struct blk_mq_ctx *ctx) | |
118 | { | |
f31967f0 JA |
119 | const int bit = ctx->index_hw[hctx->type]; |
120 | ||
121 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
122 | } |
123 | ||
f299b7c7 | 124 | struct mq_inflight { |
8446fe92 | 125 | struct block_device *part; |
a2e80f6f | 126 | unsigned int inflight[2]; |
f299b7c7 JA |
127 | }; |
128 | ||
7baa8572 | 129 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, |
f299b7c7 JA |
130 | struct request *rq, void *priv, |
131 | bool reserved) | |
132 | { | |
133 | struct mq_inflight *mi = priv; | |
134 | ||
b0d97557 JX |
135 | if ((!mi->part->bd_partno || rq->part == mi->part) && |
136 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) | |
bb4e6b14 | 137 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
138 | |
139 | return true; | |
f299b7c7 JA |
140 | } |
141 | ||
8446fe92 CH |
142 | unsigned int blk_mq_in_flight(struct request_queue *q, |
143 | struct block_device *part) | |
f299b7c7 | 144 | { |
a2e80f6f | 145 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 146 | |
f299b7c7 | 147 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 148 | |
a2e80f6f | 149 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
150 | } |
151 | ||
8446fe92 CH |
152 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
153 | unsigned int inflight[2]) | |
bf0ddaba | 154 | { |
a2e80f6f | 155 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 156 | |
bb4e6b14 | 157 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
158 | inflight[0] = mi.inflight[0]; |
159 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
160 | } |
161 | ||
1671d522 | 162 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 163 | { |
7996a8b5 BL |
164 | mutex_lock(&q->mq_freeze_lock); |
165 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 166 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 167 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 168 | if (queue_is_mq(q)) |
055f6e18 | 169 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
170 | } else { |
171 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 172 | } |
f3af020b | 173 | } |
1671d522 | 174 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 175 | |
6bae363e | 176 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 177 | { |
3ef28e83 | 178 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 179 | } |
6bae363e | 180 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 181 | |
f91328c4 KB |
182 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
183 | unsigned long timeout) | |
184 | { | |
185 | return wait_event_timeout(q->mq_freeze_wq, | |
186 | percpu_ref_is_zero(&q->q_usage_counter), | |
187 | timeout); | |
188 | } | |
189 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 190 | |
f3af020b TH |
191 | /* |
192 | * Guarantee no request is in use, so we can change any data structure of | |
193 | * the queue afterward. | |
194 | */ | |
3ef28e83 | 195 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 196 | { |
3ef28e83 DW |
197 | /* |
198 | * In the !blk_mq case we are only calling this to kill the | |
199 | * q_usage_counter, otherwise this increases the freeze depth | |
200 | * and waits for it to return to zero. For this reason there is | |
201 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
202 | * exported to drivers as the only user for unfreeze is blk_mq. | |
203 | */ | |
1671d522 | 204 | blk_freeze_queue_start(q); |
f3af020b TH |
205 | blk_mq_freeze_queue_wait(q); |
206 | } | |
3ef28e83 DW |
207 | |
208 | void blk_mq_freeze_queue(struct request_queue *q) | |
209 | { | |
210 | /* | |
211 | * ...just an alias to keep freeze and unfreeze actions balanced | |
212 | * in the blk_mq_* namespace | |
213 | */ | |
214 | blk_freeze_queue(q); | |
215 | } | |
c761d96b | 216 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 217 | |
aec89dc5 | 218 | void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) |
320ae51f | 219 | { |
7996a8b5 | 220 | mutex_lock(&q->mq_freeze_lock); |
aec89dc5 CH |
221 | if (force_atomic) |
222 | q->q_usage_counter.data->force_atomic = true; | |
7996a8b5 BL |
223 | q->mq_freeze_depth--; |
224 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
225 | if (!q->mq_freeze_depth) { | |
bdd63160 | 226 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 227 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 228 | } |
7996a8b5 | 229 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 230 | } |
aec89dc5 CH |
231 | |
232 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
233 | { | |
234 | __blk_mq_unfreeze_queue(q, false); | |
235 | } | |
b4c6a028 | 236 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 237 | |
852ec809 BVA |
238 | /* |
239 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
240 | * mpt3sas driver such that this function can be removed. | |
241 | */ | |
242 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
243 | { | |
e70feb8b ML |
244 | unsigned long flags; |
245 | ||
246 | spin_lock_irqsave(&q->queue_lock, flags); | |
247 | if (!q->quiesce_depth++) | |
248 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
249 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
852ec809 BVA |
250 | } |
251 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
252 | ||
6a83e74d | 253 | /** |
9ef4d020 | 254 | * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done |
6a83e74d BVA |
255 | * @q: request queue. |
256 | * | |
9ef4d020 ML |
257 | * Note: it is driver's responsibility for making sure that quiesce has |
258 | * been started. | |
6a83e74d | 259 | */ |
9ef4d020 | 260 | void blk_mq_wait_quiesce_done(struct request_queue *q) |
6a83e74d BVA |
261 | { |
262 | struct blk_mq_hw_ctx *hctx; | |
263 | unsigned int i; | |
264 | bool rcu = false; | |
265 | ||
6a83e74d BVA |
266 | queue_for_each_hw_ctx(q, hctx, i) { |
267 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 268 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
269 | else |
270 | rcu = true; | |
271 | } | |
272 | if (rcu) | |
273 | synchronize_rcu(); | |
274 | } | |
9ef4d020 ML |
275 | EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); |
276 | ||
277 | /** | |
278 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished | |
279 | * @q: request queue. | |
280 | * | |
281 | * Note: this function does not prevent that the struct request end_io() | |
282 | * callback function is invoked. Once this function is returned, we make | |
283 | * sure no dispatch can happen until the queue is unquiesced via | |
284 | * blk_mq_unquiesce_queue(). | |
285 | */ | |
286 | void blk_mq_quiesce_queue(struct request_queue *q) | |
287 | { | |
288 | blk_mq_quiesce_queue_nowait(q); | |
289 | blk_mq_wait_quiesce_done(q); | |
290 | } | |
6a83e74d BVA |
291 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); |
292 | ||
e4e73913 ML |
293 | /* |
294 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
295 | * @q: request queue. | |
296 | * | |
297 | * This function recovers queue into the state before quiescing | |
298 | * which is done by blk_mq_quiesce_queue. | |
299 | */ | |
300 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
301 | { | |
e70feb8b ML |
302 | unsigned long flags; |
303 | bool run_queue = false; | |
304 | ||
305 | spin_lock_irqsave(&q->queue_lock, flags); | |
306 | if (WARN_ON_ONCE(q->quiesce_depth <= 0)) { | |
307 | ; | |
308 | } else if (!--q->quiesce_depth) { | |
309 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
310 | run_queue = true; | |
311 | } | |
312 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
f4560ffe | 313 | |
1d9e9bc6 | 314 | /* dispatch requests which are inserted during quiescing */ |
e70feb8b ML |
315 | if (run_queue) |
316 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
317 | } |
318 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
319 | ||
aed3ea94 JA |
320 | void blk_mq_wake_waiters(struct request_queue *q) |
321 | { | |
322 | struct blk_mq_hw_ctx *hctx; | |
323 | unsigned int i; | |
324 | ||
325 | queue_for_each_hw_ctx(q, hctx, i) | |
326 | if (blk_mq_hw_queue_mapped(hctx)) | |
327 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
328 | } | |
329 | ||
52fdbbcc CH |
330 | void blk_rq_init(struct request_queue *q, struct request *rq) |
331 | { | |
332 | memset(rq, 0, sizeof(*rq)); | |
333 | ||
334 | INIT_LIST_HEAD(&rq->queuelist); | |
335 | rq->q = q; | |
336 | rq->__sector = (sector_t) -1; | |
337 | INIT_HLIST_NODE(&rq->hash); | |
338 | RB_CLEAR_NODE(&rq->rb_node); | |
339 | rq->tag = BLK_MQ_NO_TAG; | |
340 | rq->internal_tag = BLK_MQ_NO_TAG; | |
341 | rq->start_time_ns = ktime_get_ns(); | |
342 | rq->part = NULL; | |
343 | blk_crypto_rq_set_defaults(rq); | |
344 | } | |
345 | EXPORT_SYMBOL(blk_rq_init); | |
346 | ||
e4cdf1a1 | 347 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
fe6134f6 | 348 | struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns) |
320ae51f | 349 | { |
605f784e PB |
350 | struct blk_mq_ctx *ctx = data->ctx; |
351 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
352 | struct request_queue *q = data->q; | |
e4cdf1a1 | 353 | struct request *rq = tags->static_rqs[tag]; |
c3a148d2 | 354 | |
c7b84d42 JA |
355 | rq->q = q; |
356 | rq->mq_ctx = ctx; | |
357 | rq->mq_hctx = hctx; | |
358 | rq->cmd_flags = data->cmd_flags; | |
359 | ||
360 | if (data->flags & BLK_MQ_REQ_PM) | |
361 | data->rq_flags |= RQF_PM; | |
362 | if (blk_queue_io_stat(q)) | |
363 | data->rq_flags |= RQF_IO_STAT; | |
364 | rq->rq_flags = data->rq_flags; | |
365 | ||
56f8da64 | 366 | if (!(data->rq_flags & RQF_ELV)) { |
e4cdf1a1 | 367 | rq->tag = tag; |
76647368 | 368 | rq->internal_tag = BLK_MQ_NO_TAG; |
56f8da64 JA |
369 | } else { |
370 | rq->tag = BLK_MQ_NO_TAG; | |
371 | rq->internal_tag = tag; | |
e4cdf1a1 | 372 | } |
c7b84d42 | 373 | rq->timeout = 0; |
e4cdf1a1 | 374 | |
4f266f2b PB |
375 | if (blk_mq_need_time_stamp(rq)) |
376 | rq->start_time_ns = ktime_get_ns(); | |
377 | else | |
378 | rq->start_time_ns = 0; | |
af76e555 CH |
379 | rq->rq_disk = NULL; |
380 | rq->part = NULL; | |
6f816b4b TH |
381 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
382 | rq->alloc_time_ns = alloc_time_ns; | |
383 | #endif | |
544ccc8d | 384 | rq->io_start_time_ns = 0; |
3d244306 | 385 | rq->stats_sectors = 0; |
af76e555 CH |
386 | rq->nr_phys_segments = 0; |
387 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
388 | rq->nr_integrity_segments = 0; | |
389 | #endif | |
af76e555 CH |
390 | rq->end_io = NULL; |
391 | rq->end_io_data = NULL; | |
af76e555 | 392 | |
4f266f2b PB |
393 | blk_crypto_rq_set_defaults(rq); |
394 | INIT_LIST_HEAD(&rq->queuelist); | |
395 | /* tag was already set */ | |
396 | WRITE_ONCE(rq->deadline, 0); | |
12f5b931 | 397 | refcount_set(&rq->ref, 1); |
7ea4d8a4 | 398 | |
4f266f2b | 399 | if (rq->rq_flags & RQF_ELV) { |
7ea4d8a4 CH |
400 | struct elevator_queue *e = data->q->elevator; |
401 | ||
402 | rq->elv.icq = NULL; | |
4f266f2b PB |
403 | INIT_HLIST_NODE(&rq->hash); |
404 | RB_CLEAR_NODE(&rq->rb_node); | |
405 | ||
406 | if (!op_is_flush(data->cmd_flags) && | |
407 | e->type->ops.prepare_request) { | |
7ea4d8a4 CH |
408 | e->type->ops.prepare_request(rq); |
409 | rq->rq_flags |= RQF_ELVPRIV; | |
410 | } | |
411 | } | |
412 | ||
e4cdf1a1 | 413 | return rq; |
5dee8577 CH |
414 | } |
415 | ||
349302da JA |
416 | static inline struct request * |
417 | __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data, | |
418 | u64 alloc_time_ns) | |
419 | { | |
420 | unsigned int tag, tag_offset; | |
fe6134f6 | 421 | struct blk_mq_tags *tags; |
349302da | 422 | struct request *rq; |
fe6134f6 | 423 | unsigned long tag_mask; |
349302da JA |
424 | int i, nr = 0; |
425 | ||
fe6134f6 JA |
426 | tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset); |
427 | if (unlikely(!tag_mask)) | |
349302da JA |
428 | return NULL; |
429 | ||
fe6134f6 JA |
430 | tags = blk_mq_tags_from_data(data); |
431 | for (i = 0; tag_mask; i++) { | |
432 | if (!(tag_mask & (1UL << i))) | |
349302da JA |
433 | continue; |
434 | tag = tag_offset + i; | |
a22c00be | 435 | prefetch(tags->static_rqs[tag]); |
fe6134f6 JA |
436 | tag_mask &= ~(1UL << i); |
437 | rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns); | |
013a7f95 | 438 | rq_list_add(data->cached_rq, rq); |
c5fc7b93 | 439 | nr++; |
349302da | 440 | } |
c5fc7b93 JA |
441 | /* caller already holds a reference, add for remainder */ |
442 | percpu_ref_get_many(&data->q->q_usage_counter, nr - 1); | |
349302da JA |
443 | data->nr_tags -= nr; |
444 | ||
013a7f95 | 445 | return rq_list_pop(data->cached_rq); |
349302da JA |
446 | } |
447 | ||
b90cfaed | 448 | static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) |
d2c0d383 | 449 | { |
e6e7abff | 450 | struct request_queue *q = data->q; |
6f816b4b | 451 | u64 alloc_time_ns = 0; |
47c122e3 | 452 | struct request *rq; |
600c3b0c | 453 | unsigned int tag; |
d2c0d383 | 454 | |
6f816b4b TH |
455 | /* alloc_time includes depth and tag waits */ |
456 | if (blk_queue_rq_alloc_time(q)) | |
457 | alloc_time_ns = ktime_get_ns(); | |
458 | ||
f9afca4d | 459 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 460 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 | 461 | |
781dd830 JA |
462 | if (q->elevator) { |
463 | struct elevator_queue *e = q->elevator; | |
464 | ||
465 | data->rq_flags |= RQF_ELV; | |
466 | ||
d2c0d383 | 467 | /* |
8d663f34 | 468 | * Flush/passthrough requests are special and go directly to the |
17a51199 JA |
469 | * dispatch list. Don't include reserved tags in the |
470 | * limiting, as it isn't useful. | |
d2c0d383 | 471 | */ |
f9afca4d | 472 | if (!op_is_flush(data->cmd_flags) && |
8d663f34 | 473 | !blk_op_is_passthrough(data->cmd_flags) && |
f9afca4d | 474 | e->type->ops.limit_depth && |
17a51199 | 475 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 476 | e->type->ops.limit_depth(data->cmd_flags, data); |
d2c0d383 CH |
477 | } |
478 | ||
bf0beec0 | 479 | retry: |
600c3b0c CH |
480 | data->ctx = blk_mq_get_ctx(q); |
481 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
781dd830 | 482 | if (!(data->rq_flags & RQF_ELV)) |
600c3b0c CH |
483 | blk_mq_tag_busy(data->hctx); |
484 | ||
349302da JA |
485 | /* |
486 | * Try batched alloc if we want more than 1 tag. | |
487 | */ | |
488 | if (data->nr_tags > 1) { | |
489 | rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns); | |
490 | if (rq) | |
491 | return rq; | |
492 | data->nr_tags = 1; | |
493 | } | |
494 | ||
bf0beec0 ML |
495 | /* |
496 | * Waiting allocations only fail because of an inactive hctx. In that | |
497 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
498 | * should have migrated us to an online CPU by now. | |
499 | */ | |
e4cdf1a1 | 500 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
501 | if (tag == BLK_MQ_NO_TAG) { |
502 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
503 | return NULL; | |
bf0beec0 | 504 | /* |
349302da JA |
505 | * Give up the CPU and sleep for a random short time to |
506 | * ensure that thread using a realtime scheduling class | |
507 | * are migrated off the CPU, and thus off the hctx that | |
508 | * is going away. | |
bf0beec0 ML |
509 | */ |
510 | msleep(3); | |
511 | goto retry; | |
512 | } | |
47c122e3 | 513 | |
fe6134f6 JA |
514 | return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag, |
515 | alloc_time_ns); | |
d2c0d383 CH |
516 | } |
517 | ||
cd6ce148 | 518 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 519 | blk_mq_req_flags_t flags) |
320ae51f | 520 | { |
e6e7abff CH |
521 | struct blk_mq_alloc_data data = { |
522 | .q = q, | |
523 | .flags = flags, | |
524 | .cmd_flags = op, | |
47c122e3 | 525 | .nr_tags = 1, |
e6e7abff | 526 | }; |
bd166ef1 | 527 | struct request *rq; |
a492f075 | 528 | int ret; |
320ae51f | 529 | |
3a0a5299 | 530 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
531 | if (ret) |
532 | return ERR_PTR(ret); | |
320ae51f | 533 | |
b90cfaed | 534 | rq = __blk_mq_alloc_requests(&data); |
bd166ef1 | 535 | if (!rq) |
a5ea5811 | 536 | goto out_queue_exit; |
0c4de0f3 CH |
537 | rq->__data_len = 0; |
538 | rq->__sector = (sector_t) -1; | |
539 | rq->bio = rq->biotail = NULL; | |
320ae51f | 540 | return rq; |
a5ea5811 CH |
541 | out_queue_exit: |
542 | blk_queue_exit(q); | |
543 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 544 | } |
4bb659b1 | 545 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 546 | |
cd6ce148 | 547 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 548 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 549 | { |
e6e7abff CH |
550 | struct blk_mq_alloc_data data = { |
551 | .q = q, | |
552 | .flags = flags, | |
553 | .cmd_flags = op, | |
47c122e3 | 554 | .nr_tags = 1, |
e6e7abff | 555 | }; |
600c3b0c | 556 | u64 alloc_time_ns = 0; |
6d2809d5 | 557 | unsigned int cpu; |
600c3b0c | 558 | unsigned int tag; |
1f5bd336 ML |
559 | int ret; |
560 | ||
600c3b0c CH |
561 | /* alloc_time includes depth and tag waits */ |
562 | if (blk_queue_rq_alloc_time(q)) | |
563 | alloc_time_ns = ktime_get_ns(); | |
564 | ||
1f5bd336 ML |
565 | /* |
566 | * If the tag allocator sleeps we could get an allocation for a | |
567 | * different hardware context. No need to complicate the low level | |
568 | * allocator for this for the rare use case of a command tied to | |
569 | * a specific queue. | |
570 | */ | |
600c3b0c | 571 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) |
1f5bd336 ML |
572 | return ERR_PTR(-EINVAL); |
573 | ||
574 | if (hctx_idx >= q->nr_hw_queues) | |
575 | return ERR_PTR(-EIO); | |
576 | ||
3a0a5299 | 577 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
578 | if (ret) |
579 | return ERR_PTR(ret); | |
580 | ||
c8712c6a CH |
581 | /* |
582 | * Check if the hardware context is actually mapped to anything. | |
583 | * If not tell the caller that it should skip this queue. | |
584 | */ | |
a5ea5811 | 585 | ret = -EXDEV; |
e6e7abff CH |
586 | data.hctx = q->queue_hw_ctx[hctx_idx]; |
587 | if (!blk_mq_hw_queue_mapped(data.hctx)) | |
a5ea5811 | 588 | goto out_queue_exit; |
e6e7abff CH |
589 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
590 | data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 591 | |
42fdc5e4 | 592 | if (!q->elevator) |
600c3b0c | 593 | blk_mq_tag_busy(data.hctx); |
781dd830 JA |
594 | else |
595 | data.rq_flags |= RQF_ELV; | |
600c3b0c | 596 | |
a5ea5811 | 597 | ret = -EWOULDBLOCK; |
600c3b0c CH |
598 | tag = blk_mq_get_tag(&data); |
599 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 600 | goto out_queue_exit; |
fe6134f6 JA |
601 | return blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag, |
602 | alloc_time_ns); | |
600c3b0c | 603 | |
a5ea5811 CH |
604 | out_queue_exit: |
605 | blk_queue_exit(q); | |
606 | return ERR_PTR(ret); | |
1f5bd336 ML |
607 | } |
608 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
609 | ||
12f5b931 KB |
610 | static void __blk_mq_free_request(struct request *rq) |
611 | { | |
612 | struct request_queue *q = rq->q; | |
613 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 614 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
615 | const int sched_tag = rq->internal_tag; |
616 | ||
a892c8d5 | 617 | blk_crypto_free_request(rq); |
986d413b | 618 | blk_pm_mark_last_busy(rq); |
ea4f995e | 619 | rq->mq_hctx = NULL; |
76647368 | 620 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 621 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 622 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 623 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
624 | blk_mq_sched_restart(hctx); |
625 | blk_queue_exit(q); | |
626 | } | |
627 | ||
6af54051 | 628 | void blk_mq_free_request(struct request *rq) |
320ae51f | 629 | { |
320ae51f | 630 | struct request_queue *q = rq->q; |
ea4f995e | 631 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 632 | |
5bbf4e5a | 633 | if (rq->rq_flags & RQF_ELVPRIV) { |
2ff0682d JA |
634 | struct elevator_queue *e = q->elevator; |
635 | ||
636 | if (e->type->ops.finish_request) | |
f9cd4bfe | 637 | e->type->ops.finish_request(rq); |
6af54051 CH |
638 | if (rq->elv.icq) { |
639 | put_io_context(rq->elv.icq->ioc); | |
640 | rq->elv.icq = NULL; | |
641 | } | |
642 | } | |
320ae51f | 643 | |
e8064021 | 644 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 645 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 646 | |
7beb2f84 | 647 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 648 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 649 | |
a7905043 | 650 | rq_qos_done(q, rq); |
0d2602ca | 651 | |
12f5b931 KB |
652 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
653 | if (refcount_dec_and_test(&rq->ref)) | |
654 | __blk_mq_free_request(rq); | |
320ae51f | 655 | } |
1a3b595a | 656 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 657 | |
47c122e3 | 658 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
320ae51f | 659 | { |
013a7f95 | 660 | struct request *rq; |
fe1f4526 | 661 | |
c5fc7b93 | 662 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) |
47c122e3 | 663 | blk_mq_free_request(rq); |
47c122e3 | 664 | } |
522a7775 | 665 | |
22350ad7 CH |
666 | void blk_dump_rq_flags(struct request *rq, char *msg) |
667 | { | |
668 | printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, | |
669 | rq->rq_disk ? rq->rq_disk->disk_name : "?", | |
670 | (unsigned long long) rq->cmd_flags); | |
671 | ||
672 | printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", | |
673 | (unsigned long long)blk_rq_pos(rq), | |
674 | blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); | |
675 | printk(KERN_INFO " bio %p, biotail %p, len %u\n", | |
676 | rq->bio, rq->biotail, blk_rq_bytes(rq)); | |
677 | } | |
678 | EXPORT_SYMBOL(blk_dump_rq_flags); | |
679 | ||
9be3e06f JA |
680 | static void req_bio_endio(struct request *rq, struct bio *bio, |
681 | unsigned int nbytes, blk_status_t error) | |
682 | { | |
478eb72b | 683 | if (unlikely(error)) { |
9be3e06f | 684 | bio->bi_status = error; |
478eb72b | 685 | } else if (req_op(rq) == REQ_OP_ZONE_APPEND) { |
9be3e06f JA |
686 | /* |
687 | * Partial zone append completions cannot be supported as the | |
688 | * BIO fragments may end up not being written sequentially. | |
689 | */ | |
297db731 | 690 | if (bio->bi_iter.bi_size != nbytes) |
9be3e06f JA |
691 | bio->bi_status = BLK_STS_IOERR; |
692 | else | |
693 | bio->bi_iter.bi_sector = rq->__sector; | |
694 | } | |
695 | ||
478eb72b PB |
696 | bio_advance(bio, nbytes); |
697 | ||
698 | if (unlikely(rq->rq_flags & RQF_QUIET)) | |
699 | bio_set_flag(bio, BIO_QUIET); | |
9be3e06f JA |
700 | /* don't actually finish bio if it's part of flush sequence */ |
701 | if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) | |
702 | bio_endio(bio); | |
703 | } | |
704 | ||
705 | static void blk_account_io_completion(struct request *req, unsigned int bytes) | |
706 | { | |
707 | if (req->part && blk_do_io_stat(req)) { | |
708 | const int sgrp = op_stat_group(req_op(req)); | |
709 | ||
710 | part_stat_lock(); | |
711 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); | |
712 | part_stat_unlock(); | |
713 | } | |
714 | } | |
715 | ||
0d7a29a2 CH |
716 | static void blk_print_req_error(struct request *req, blk_status_t status) |
717 | { | |
718 | printk_ratelimited(KERN_ERR | |
719 | "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " | |
720 | "phys_seg %u prio class %u\n", | |
721 | blk_status_to_str(status), | |
722 | req->rq_disk ? req->rq_disk->disk_name : "?", | |
723 | blk_rq_pos(req), req_op(req), blk_op_str(req_op(req)), | |
724 | req->cmd_flags & ~REQ_OP_MASK, | |
725 | req->nr_phys_segments, | |
726 | IOPRIO_PRIO_CLASS(req->ioprio)); | |
727 | } | |
728 | ||
9be3e06f JA |
729 | /** |
730 | * blk_update_request - Complete multiple bytes without completing the request | |
731 | * @req: the request being processed | |
732 | * @error: block status code | |
733 | * @nr_bytes: number of bytes to complete for @req | |
734 | * | |
735 | * Description: | |
736 | * Ends I/O on a number of bytes attached to @req, but doesn't complete | |
737 | * the request structure even if @req doesn't have leftover. | |
738 | * If @req has leftover, sets it up for the next range of segments. | |
739 | * | |
740 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees | |
741 | * %false return from this function. | |
742 | * | |
743 | * Note: | |
744 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function | |
745 | * except in the consistency check at the end of this function. | |
746 | * | |
747 | * Return: | |
748 | * %false - this request doesn't have any more data | |
749 | * %true - this request has more data | |
750 | **/ | |
751 | bool blk_update_request(struct request *req, blk_status_t error, | |
752 | unsigned int nr_bytes) | |
753 | { | |
754 | int total_bytes; | |
755 | ||
8a7d267b | 756 | trace_block_rq_complete(req, error, nr_bytes); |
9be3e06f JA |
757 | |
758 | if (!req->bio) | |
759 | return false; | |
760 | ||
761 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
762 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && | |
763 | error == BLK_STS_OK) | |
764 | req->q->integrity.profile->complete_fn(req, nr_bytes); | |
765 | #endif | |
766 | ||
767 | if (unlikely(error && !blk_rq_is_passthrough(req) && | |
768 | !(req->rq_flags & RQF_QUIET))) | |
769 | blk_print_req_error(req, error); | |
770 | ||
771 | blk_account_io_completion(req, nr_bytes); | |
772 | ||
773 | total_bytes = 0; | |
774 | while (req->bio) { | |
775 | struct bio *bio = req->bio; | |
776 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); | |
777 | ||
778 | if (bio_bytes == bio->bi_iter.bi_size) | |
779 | req->bio = bio->bi_next; | |
780 | ||
781 | /* Completion has already been traced */ | |
782 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
783 | req_bio_endio(req, bio, bio_bytes, error); | |
784 | ||
785 | total_bytes += bio_bytes; | |
786 | nr_bytes -= bio_bytes; | |
787 | ||
788 | if (!nr_bytes) | |
789 | break; | |
790 | } | |
791 | ||
792 | /* | |
793 | * completely done | |
794 | */ | |
795 | if (!req->bio) { | |
796 | /* | |
797 | * Reset counters so that the request stacking driver | |
798 | * can find how many bytes remain in the request | |
799 | * later. | |
800 | */ | |
801 | req->__data_len = 0; | |
802 | return false; | |
803 | } | |
804 | ||
805 | req->__data_len -= total_bytes; | |
806 | ||
807 | /* update sector only for requests with clear definition of sector */ | |
808 | if (!blk_rq_is_passthrough(req)) | |
809 | req->__sector += total_bytes >> 9; | |
810 | ||
811 | /* mixed attributes always follow the first bio */ | |
812 | if (req->rq_flags & RQF_MIXED_MERGE) { | |
813 | req->cmd_flags &= ~REQ_FAILFAST_MASK; | |
814 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; | |
815 | } | |
816 | ||
817 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { | |
818 | /* | |
819 | * If total number of sectors is less than the first segment | |
820 | * size, something has gone terribly wrong. | |
821 | */ | |
822 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { | |
823 | blk_dump_rq_flags(req, "request botched"); | |
824 | req->__data_len = blk_rq_cur_bytes(req); | |
825 | } | |
826 | ||
827 | /* recalculate the number of segments */ | |
828 | req->nr_phys_segments = blk_recalc_rq_segments(req); | |
829 | } | |
830 | ||
831 | return true; | |
832 | } | |
833 | EXPORT_SYMBOL_GPL(blk_update_request); | |
834 | ||
450b7879 CH |
835 | static void __blk_account_io_done(struct request *req, u64 now) |
836 | { | |
837 | const int sgrp = op_stat_group(req_op(req)); | |
838 | ||
839 | part_stat_lock(); | |
840 | update_io_ticks(req->part, jiffies, true); | |
841 | part_stat_inc(req->part, ios[sgrp]); | |
842 | part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); | |
843 | part_stat_unlock(); | |
844 | } | |
845 | ||
846 | static inline void blk_account_io_done(struct request *req, u64 now) | |
847 | { | |
848 | /* | |
849 | * Account IO completion. flush_rq isn't accounted as a | |
850 | * normal IO on queueing nor completion. Accounting the | |
851 | * containing request is enough. | |
852 | */ | |
853 | if (blk_do_io_stat(req) && req->part && | |
854 | !(req->rq_flags & RQF_FLUSH_SEQ)) | |
855 | __blk_account_io_done(req, now); | |
856 | } | |
857 | ||
858 | static void __blk_account_io_start(struct request *rq) | |
859 | { | |
860 | /* passthrough requests can hold bios that do not have ->bi_bdev set */ | |
861 | if (rq->bio && rq->bio->bi_bdev) | |
862 | rq->part = rq->bio->bi_bdev; | |
863 | else | |
864 | rq->part = rq->rq_disk->part0; | |
865 | ||
866 | part_stat_lock(); | |
867 | update_io_ticks(rq->part, jiffies, false); | |
868 | part_stat_unlock(); | |
869 | } | |
870 | ||
871 | static inline void blk_account_io_start(struct request *req) | |
872 | { | |
873 | if (blk_do_io_stat(req)) | |
874 | __blk_account_io_start(req); | |
875 | } | |
876 | ||
f794f335 | 877 | static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) |
320ae51f | 878 | { |
4bc6339a OS |
879 | if (rq->rq_flags & RQF_STATS) { |
880 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 881 | blk_stat_add(rq, now); |
4bc6339a OS |
882 | } |
883 | ||
87890092 | 884 | blk_mq_sched_completed_request(rq, now); |
522a7775 | 885 | blk_account_io_done(rq, now); |
f794f335 | 886 | } |
522a7775 | 887 | |
f794f335 JA |
888 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
889 | { | |
890 | if (blk_mq_need_time_stamp(rq)) | |
891 | __blk_mq_end_request_acct(rq, ktime_get_ns()); | |
0d11e6ac | 892 | |
91b63639 | 893 | if (rq->end_io) { |
a7905043 | 894 | rq_qos_done(rq->q, rq); |
320ae51f | 895 | rq->end_io(rq, error); |
91b63639 | 896 | } else { |
320ae51f | 897 | blk_mq_free_request(rq); |
91b63639 | 898 | } |
320ae51f | 899 | } |
c8a446ad | 900 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 901 | |
2a842aca | 902 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
903 | { |
904 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
905 | BUG(); | |
c8a446ad | 906 | __blk_mq_end_request(rq, error); |
63151a44 | 907 | } |
c8a446ad | 908 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 909 | |
f794f335 JA |
910 | #define TAG_COMP_BATCH 32 |
911 | ||
912 | static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, | |
913 | int *tag_array, int nr_tags) | |
914 | { | |
915 | struct request_queue *q = hctx->queue; | |
916 | ||
3b87c6ea ML |
917 | /* |
918 | * All requests should have been marked as RQF_MQ_INFLIGHT, so | |
919 | * update hctx->nr_active in batch | |
920 | */ | |
921 | if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) | |
922 | __blk_mq_sub_active_requests(hctx, nr_tags); | |
923 | ||
f794f335 JA |
924 | blk_mq_put_tags(hctx->tags, tag_array, nr_tags); |
925 | percpu_ref_put_many(&q->q_usage_counter, nr_tags); | |
926 | } | |
927 | ||
928 | void blk_mq_end_request_batch(struct io_comp_batch *iob) | |
929 | { | |
930 | int tags[TAG_COMP_BATCH], nr_tags = 0; | |
02f7eab0 | 931 | struct blk_mq_hw_ctx *cur_hctx = NULL; |
f794f335 JA |
932 | struct request *rq; |
933 | u64 now = 0; | |
934 | ||
935 | if (iob->need_ts) | |
936 | now = ktime_get_ns(); | |
937 | ||
938 | while ((rq = rq_list_pop(&iob->req_list)) != NULL) { | |
939 | prefetch(rq->bio); | |
940 | prefetch(rq->rq_next); | |
941 | ||
942 | blk_update_request(rq, BLK_STS_OK, blk_rq_bytes(rq)); | |
943 | if (iob->need_ts) | |
944 | __blk_mq_end_request_acct(rq, now); | |
945 | ||
98b26a0e JA |
946 | rq_qos_done(rq->q, rq); |
947 | ||
f794f335 JA |
948 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
949 | if (!refcount_dec_and_test(&rq->ref)) | |
950 | continue; | |
951 | ||
952 | blk_crypto_free_request(rq); | |
953 | blk_pm_mark_last_busy(rq); | |
f794f335 | 954 | |
02f7eab0 JA |
955 | if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { |
956 | if (cur_hctx) | |
957 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); | |
f794f335 | 958 | nr_tags = 0; |
02f7eab0 | 959 | cur_hctx = rq->mq_hctx; |
f794f335 JA |
960 | } |
961 | tags[nr_tags++] = rq->tag; | |
f794f335 JA |
962 | } |
963 | ||
964 | if (nr_tags) | |
02f7eab0 | 965 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
f794f335 JA |
966 | } |
967 | EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); | |
968 | ||
f9ab4918 | 969 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 970 | { |
f9ab4918 SAS |
971 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
972 | struct request *rq, *next; | |
c3077b5d | 973 | |
f9ab4918 | 974 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 975 | rq->q->mq_ops->complete(rq); |
320ae51f | 976 | } |
320ae51f | 977 | |
f9ab4918 | 978 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 979 | { |
f9ab4918 | 980 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
981 | } |
982 | ||
c3077b5d CH |
983 | static int blk_softirq_cpu_dead(unsigned int cpu) |
984 | { | |
f9ab4918 | 985 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
986 | return 0; |
987 | } | |
988 | ||
40d09b53 | 989 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 990 | { |
f9ab4918 | 991 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
992 | } |
993 | ||
96339526 CH |
994 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
995 | { | |
996 | int cpu = raw_smp_processor_id(); | |
997 | ||
998 | if (!IS_ENABLED(CONFIG_SMP) || | |
999 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
1000 | return false; | |
71425189 SAS |
1001 | /* |
1002 | * With force threaded interrupts enabled, raising softirq from an SMP | |
1003 | * function call will always result in waking the ksoftirqd thread. | |
1004 | * This is probably worse than completing the request on a different | |
1005 | * cache domain. | |
1006 | */ | |
91cc470e | 1007 | if (force_irqthreads()) |
71425189 | 1008 | return false; |
96339526 CH |
1009 | |
1010 | /* same CPU or cache domain? Complete locally */ | |
1011 | if (cpu == rq->mq_ctx->cpu || | |
1012 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
1013 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
1014 | return false; | |
1015 | ||
1016 | /* don't try to IPI to an offline CPU */ | |
1017 | return cpu_online(rq->mq_ctx->cpu); | |
1018 | } | |
1019 | ||
f9ab4918 SAS |
1020 | static void blk_mq_complete_send_ipi(struct request *rq) |
1021 | { | |
1022 | struct llist_head *list; | |
1023 | unsigned int cpu; | |
1024 | ||
1025 | cpu = rq->mq_ctx->cpu; | |
1026 | list = &per_cpu(blk_cpu_done, cpu); | |
1027 | if (llist_add(&rq->ipi_list, list)) { | |
1028 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); | |
1029 | smp_call_function_single_async(cpu, &rq->csd); | |
1030 | } | |
1031 | } | |
1032 | ||
1033 | static void blk_mq_raise_softirq(struct request *rq) | |
1034 | { | |
1035 | struct llist_head *list; | |
1036 | ||
1037 | preempt_disable(); | |
1038 | list = this_cpu_ptr(&blk_cpu_done); | |
1039 | if (llist_add(&rq->ipi_list, list)) | |
1040 | raise_softirq(BLOCK_SOFTIRQ); | |
1041 | preempt_enable(); | |
1042 | } | |
1043 | ||
40d09b53 | 1044 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 1045 | { |
af78ff7c | 1046 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 1047 | |
4ab32bf3 JA |
1048 | /* |
1049 | * For a polled request, always complete locallly, it's pointless | |
1050 | * to redirect the completion. | |
1051 | */ | |
6ce913fe | 1052 | if (rq->cmd_flags & REQ_POLLED) |
40d09b53 | 1053 | return false; |
38535201 | 1054 | |
96339526 | 1055 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
1056 | blk_mq_complete_send_ipi(rq); |
1057 | return true; | |
3d6efbf6 | 1058 | } |
40d09b53 | 1059 | |
f9ab4918 SAS |
1060 | if (rq->q->nr_hw_queues == 1) { |
1061 | blk_mq_raise_softirq(rq); | |
1062 | return true; | |
1063 | } | |
1064 | return false; | |
40d09b53 CH |
1065 | } |
1066 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
1067 | ||
1068 | /** | |
1069 | * blk_mq_complete_request - end I/O on a request | |
1070 | * @rq: the request being processed | |
1071 | * | |
1072 | * Description: | |
1073 | * Complete a request by scheduling the ->complete_rq operation. | |
1074 | **/ | |
1075 | void blk_mq_complete_request(struct request *rq) | |
1076 | { | |
1077 | if (!blk_mq_complete_request_remote(rq)) | |
1078 | rq->q->mq_ops->complete(rq); | |
320ae51f | 1079 | } |
15f73f5b | 1080 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 1081 | |
04ced159 | 1082 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 1083 | __releases(hctx->srcu) |
04ced159 JA |
1084 | { |
1085 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
1086 | rcu_read_unlock(); | |
1087 | else | |
05707b64 | 1088 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
1089 | } |
1090 | ||
1091 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 1092 | __acquires(hctx->srcu) |
04ced159 | 1093 | { |
08b5a6e2 JA |
1094 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
1095 | /* shut up gcc false positive */ | |
1096 | *srcu_idx = 0; | |
04ced159 | 1097 | rcu_read_lock(); |
08b5a6e2 | 1098 | } else |
05707b64 | 1099 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
1100 | } |
1101 | ||
105663f7 AA |
1102 | /** |
1103 | * blk_mq_start_request - Start processing a request | |
1104 | * @rq: Pointer to request to be started | |
1105 | * | |
1106 | * Function used by device drivers to notify the block layer that a request | |
1107 | * is going to be processed now, so blk layer can do proper initializations | |
1108 | * such as starting the timeout timer. | |
1109 | */ | |
e2490073 | 1110 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
1111 | { |
1112 | struct request_queue *q = rq->q; | |
1113 | ||
a54895fa | 1114 | trace_block_rq_issue(rq); |
320ae51f | 1115 | |
cf43e6be | 1116 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
00067077 JA |
1117 | u64 start_time; |
1118 | #ifdef CONFIG_BLK_CGROUP | |
1119 | if (rq->bio) | |
1120 | start_time = bio_issue_time(&rq->bio->bi_issue); | |
1121 | else | |
1122 | #endif | |
1123 | start_time = ktime_get_ns(); | |
1124 | rq->io_start_time_ns = start_time; | |
3d244306 | 1125 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 1126 | rq->rq_flags |= RQF_STATS; |
a7905043 | 1127 | rq_qos_issue(q, rq); |
cf43e6be JA |
1128 | } |
1129 | ||
1d9bd516 | 1130 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 1131 | |
1d9bd516 | 1132 | blk_add_timer(rq); |
12f5b931 | 1133 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 1134 | |
54d4e6ab MG |
1135 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
1136 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
1137 | q->integrity.profile->prepare_fn(rq); | |
1138 | #endif | |
3e08773c CH |
1139 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
1140 | WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq)); | |
320ae51f | 1141 | } |
e2490073 | 1142 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 1143 | |
4054cff9 CH |
1144 | /** |
1145 | * blk_end_sync_rq - executes a completion event on a request | |
1146 | * @rq: request to complete | |
1147 | * @error: end I/O status of the request | |
1148 | */ | |
1149 | static void blk_end_sync_rq(struct request *rq, blk_status_t error) | |
1150 | { | |
1151 | struct completion *waiting = rq->end_io_data; | |
1152 | ||
1153 | rq->end_io_data = (void *)(uintptr_t)error; | |
1154 | ||
1155 | /* | |
1156 | * complete last, if this is a stack request the process (and thus | |
1157 | * the rq pointer) could be invalid right after this complete() | |
1158 | */ | |
1159 | complete(waiting); | |
1160 | } | |
1161 | ||
1162 | /** | |
1163 | * blk_execute_rq_nowait - insert a request to I/O scheduler for execution | |
1164 | * @bd_disk: matching gendisk | |
1165 | * @rq: request to insert | |
1166 | * @at_head: insert request at head or tail of queue | |
1167 | * @done: I/O completion handler | |
1168 | * | |
1169 | * Description: | |
1170 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1171 | * for execution. Don't wait for completion. | |
1172 | * | |
1173 | * Note: | |
1174 | * This function will invoke @done directly if the queue is dead. | |
1175 | */ | |
1176 | void blk_execute_rq_nowait(struct gendisk *bd_disk, struct request *rq, | |
1177 | int at_head, rq_end_io_fn *done) | |
1178 | { | |
1179 | WARN_ON(irqs_disabled()); | |
1180 | WARN_ON(!blk_rq_is_passthrough(rq)); | |
1181 | ||
1182 | rq->rq_disk = bd_disk; | |
1183 | rq->end_io = done; | |
1184 | ||
1185 | blk_account_io_start(rq); | |
1186 | ||
1187 | /* | |
1188 | * don't check dying flag for MQ because the request won't | |
1189 | * be reused after dying flag is set | |
1190 | */ | |
1191 | blk_mq_sched_insert_request(rq, at_head, true, false); | |
1192 | } | |
1193 | EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); | |
1194 | ||
1195 | static bool blk_rq_is_poll(struct request *rq) | |
1196 | { | |
1197 | if (!rq->mq_hctx) | |
1198 | return false; | |
1199 | if (rq->mq_hctx->type != HCTX_TYPE_POLL) | |
1200 | return false; | |
1201 | if (WARN_ON_ONCE(!rq->bio)) | |
1202 | return false; | |
1203 | return true; | |
1204 | } | |
1205 | ||
1206 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) | |
1207 | { | |
1208 | do { | |
1209 | bio_poll(rq->bio, NULL, 0); | |
1210 | cond_resched(); | |
1211 | } while (!completion_done(wait)); | |
1212 | } | |
1213 | ||
1214 | /** | |
1215 | * blk_execute_rq - insert a request into queue for execution | |
1216 | * @bd_disk: matching gendisk | |
1217 | * @rq: request to insert | |
1218 | * @at_head: insert request at head or tail of queue | |
1219 | * | |
1220 | * Description: | |
1221 | * Insert a fully prepared request at the back of the I/O scheduler queue | |
1222 | * for execution and wait for completion. | |
1223 | * Return: The blk_status_t result provided to blk_mq_end_request(). | |
1224 | */ | |
1225 | blk_status_t blk_execute_rq(struct gendisk *bd_disk, struct request *rq, | |
1226 | int at_head) | |
1227 | { | |
1228 | DECLARE_COMPLETION_ONSTACK(wait); | |
1229 | unsigned long hang_check; | |
1230 | ||
1231 | rq->end_io_data = &wait; | |
1232 | blk_execute_rq_nowait(bd_disk, rq, at_head, blk_end_sync_rq); | |
1233 | ||
1234 | /* Prevent hang_check timer from firing at us during very long I/O */ | |
1235 | hang_check = sysctl_hung_task_timeout_secs; | |
1236 | ||
1237 | if (blk_rq_is_poll(rq)) | |
1238 | blk_rq_poll_completion(rq, &wait); | |
1239 | else if (hang_check) | |
1240 | while (!wait_for_completion_io_timeout(&wait, | |
1241 | hang_check * (HZ/2))) | |
1242 | ; | |
1243 | else | |
1244 | wait_for_completion_io(&wait); | |
1245 | ||
1246 | return (blk_status_t)(uintptr_t)rq->end_io_data; | |
1247 | } | |
1248 | EXPORT_SYMBOL(blk_execute_rq); | |
1249 | ||
ed0791b2 | 1250 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
1251 | { |
1252 | struct request_queue *q = rq->q; | |
1253 | ||
923218f6 ML |
1254 | blk_mq_put_driver_tag(rq); |
1255 | ||
a54895fa | 1256 | trace_block_rq_requeue(rq); |
a7905043 | 1257 | rq_qos_requeue(q, rq); |
49f5baa5 | 1258 | |
12f5b931 KB |
1259 | if (blk_mq_request_started(rq)) { |
1260 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 1261 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 1262 | } |
320ae51f JA |
1263 | } |
1264 | ||
2b053aca | 1265 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 1266 | { |
ed0791b2 | 1267 | __blk_mq_requeue_request(rq); |
ed0791b2 | 1268 | |
105976f5 ML |
1269 | /* this request will be re-inserted to io scheduler queue */ |
1270 | blk_mq_sched_requeue_request(rq); | |
1271 | ||
2b053aca | 1272 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
1273 | } |
1274 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
1275 | ||
6fca6a61 CH |
1276 | static void blk_mq_requeue_work(struct work_struct *work) |
1277 | { | |
1278 | struct request_queue *q = | |
2849450a | 1279 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
1280 | LIST_HEAD(rq_list); |
1281 | struct request *rq, *next; | |
6fca6a61 | 1282 | |
18e9781d | 1283 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 1284 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 1285 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
1286 | |
1287 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 1288 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
1289 | continue; |
1290 | ||
e8064021 | 1291 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 1292 | list_del_init(&rq->queuelist); |
aef1897c JW |
1293 | /* |
1294 | * If RQF_DONTPREP, rq has contained some driver specific | |
1295 | * data, so insert it to hctx dispatch list to avoid any | |
1296 | * merge. | |
1297 | */ | |
1298 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 1299 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
1300 | else |
1301 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
1302 | } |
1303 | ||
1304 | while (!list_empty(&rq_list)) { | |
1305 | rq = list_entry(rq_list.next, struct request, queuelist); | |
1306 | list_del_init(&rq->queuelist); | |
9e97d295 | 1307 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
1308 | } |
1309 | ||
52d7f1b5 | 1310 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
1311 | } |
1312 | ||
2b053aca BVA |
1313 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
1314 | bool kick_requeue_list) | |
6fca6a61 CH |
1315 | { |
1316 | struct request_queue *q = rq->q; | |
1317 | unsigned long flags; | |
1318 | ||
1319 | /* | |
1320 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 1321 | * request head insertion from the workqueue. |
6fca6a61 | 1322 | */ |
e8064021 | 1323 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
1324 | |
1325 | spin_lock_irqsave(&q->requeue_lock, flags); | |
1326 | if (at_head) { | |
e8064021 | 1327 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
1328 | list_add(&rq->queuelist, &q->requeue_list); |
1329 | } else { | |
1330 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
1331 | } | |
1332 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
1333 | |
1334 | if (kick_requeue_list) | |
1335 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 1336 | } |
6fca6a61 CH |
1337 | |
1338 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
1339 | { | |
ae943d20 | 1340 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
1341 | } |
1342 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
1343 | ||
2849450a MS |
1344 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
1345 | unsigned long msecs) | |
1346 | { | |
d4acf365 BVA |
1347 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
1348 | msecs_to_jiffies(msecs)); | |
2849450a MS |
1349 | } |
1350 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
1351 | ||
3c94d83c JA |
1352 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1353 | void *priv, bool reserved) | |
ae879912 JA |
1354 | { |
1355 | /* | |
05a4fed6 | 1356 | * If we find a request that isn't idle and the queue matches, |
3c94d83c | 1357 | * we know the queue is busy. Return false to stop the iteration. |
ae879912 | 1358 | */ |
05a4fed6 | 1359 | if (blk_mq_request_started(rq) && rq->q == hctx->queue) { |
ae879912 JA |
1360 | bool *busy = priv; |
1361 | ||
1362 | *busy = true; | |
1363 | return false; | |
1364 | } | |
1365 | ||
1366 | return true; | |
1367 | } | |
1368 | ||
3c94d83c | 1369 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
1370 | { |
1371 | bool busy = false; | |
1372 | ||
3c94d83c | 1373 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
1374 | return busy; |
1375 | } | |
3c94d83c | 1376 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 1377 | |
358f70da | 1378 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 1379 | { |
da661267 | 1380 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
1381 | if (req->q->mq_ops->timeout) { |
1382 | enum blk_eh_timer_return ret; | |
1383 | ||
1384 | ret = req->q->mq_ops->timeout(req, reserved); | |
1385 | if (ret == BLK_EH_DONE) | |
1386 | return; | |
1387 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 1388 | } |
d1210d5a CH |
1389 | |
1390 | blk_add_timer(req); | |
87ee7b11 | 1391 | } |
5b3f25fc | 1392 | |
12f5b931 | 1393 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 1394 | { |
12f5b931 | 1395 | unsigned long deadline; |
87ee7b11 | 1396 | |
12f5b931 KB |
1397 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
1398 | return false; | |
da661267 CH |
1399 | if (rq->rq_flags & RQF_TIMED_OUT) |
1400 | return false; | |
a7af0af3 | 1401 | |
079076b3 | 1402 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
1403 | if (time_after_eq(jiffies, deadline)) |
1404 | return true; | |
a7af0af3 | 1405 | |
12f5b931 KB |
1406 | if (*next == 0) |
1407 | *next = deadline; | |
1408 | else if (time_after(*next, deadline)) | |
1409 | *next = deadline; | |
1410 | return false; | |
87ee7b11 JA |
1411 | } |
1412 | ||
2e315dc0 ML |
1413 | void blk_mq_put_rq_ref(struct request *rq) |
1414 | { | |
a9ed27a7 | 1415 | if (is_flush_rq(rq)) |
2e315dc0 ML |
1416 | rq->end_io(rq, 0); |
1417 | else if (refcount_dec_and_test(&rq->ref)) | |
1418 | __blk_mq_free_request(rq); | |
1419 | } | |
1420 | ||
7baa8572 | 1421 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
1422 | struct request *rq, void *priv, bool reserved) |
1423 | { | |
12f5b931 KB |
1424 | unsigned long *next = priv; |
1425 | ||
1426 | /* | |
c797b40c ML |
1427 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
1428 | * be reallocated underneath the timeout handler's processing, then | |
1429 | * the expire check is reliable. If the request is not expired, then | |
1430 | * it was completed and reallocated as a new request after returning | |
1431 | * from blk_mq_check_expired(). | |
1d9bd516 | 1432 | */ |
12f5b931 | 1433 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 1434 | blk_mq_rq_timed_out(rq, reserved); |
7baa8572 | 1435 | return true; |
1d9bd516 TH |
1436 | } |
1437 | ||
287922eb | 1438 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 1439 | { |
287922eb CH |
1440 | struct request_queue *q = |
1441 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 1442 | unsigned long next = 0; |
1d9bd516 | 1443 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 1444 | int i; |
320ae51f | 1445 | |
71f79fb3 GKB |
1446 | /* A deadlock might occur if a request is stuck requiring a |
1447 | * timeout at the same time a queue freeze is waiting | |
1448 | * completion, since the timeout code would not be able to | |
1449 | * acquire the queue reference here. | |
1450 | * | |
1451 | * That's why we don't use blk_queue_enter here; instead, we use | |
1452 | * percpu_ref_tryget directly, because we need to be able to | |
1453 | * obtain a reference even in the short window between the queue | |
1454 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1455 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1456 | * consumed, marked by the instant q_usage_counter reaches |
1457 | * zero. | |
1458 | */ | |
1459 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1460 | return; |
1461 | ||
12f5b931 | 1462 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 1463 | |
12f5b931 KB |
1464 | if (next != 0) { |
1465 | mod_timer(&q->timeout, next); | |
0d2602ca | 1466 | } else { |
fcd36c36 BVA |
1467 | /* |
1468 | * Request timeouts are handled as a forward rolling timer. If | |
1469 | * we end up here it means that no requests are pending and | |
1470 | * also that no request has been pending for a while. Mark | |
1471 | * each hctx as idle. | |
1472 | */ | |
f054b56c ML |
1473 | queue_for_each_hw_ctx(q, hctx, i) { |
1474 | /* the hctx may be unmapped, so check it here */ | |
1475 | if (blk_mq_hw_queue_mapped(hctx)) | |
1476 | blk_mq_tag_idle(hctx); | |
1477 | } | |
0d2602ca | 1478 | } |
287922eb | 1479 | blk_queue_exit(q); |
320ae51f JA |
1480 | } |
1481 | ||
88459642 OS |
1482 | struct flush_busy_ctx_data { |
1483 | struct blk_mq_hw_ctx *hctx; | |
1484 | struct list_head *list; | |
1485 | }; | |
1486 | ||
1487 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1488 | { | |
1489 | struct flush_busy_ctx_data *flush_data = data; | |
1490 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1491 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1492 | enum hctx_type type = hctx->type; |
88459642 | 1493 | |
88459642 | 1494 | spin_lock(&ctx->lock); |
c16d6b5a | 1495 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1496 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1497 | spin_unlock(&ctx->lock); |
1498 | return true; | |
1499 | } | |
1500 | ||
1429d7c9 JA |
1501 | /* |
1502 | * Process software queues that have been marked busy, splicing them | |
1503 | * to the for-dispatch | |
1504 | */ | |
2c3ad667 | 1505 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1506 | { |
88459642 OS |
1507 | struct flush_busy_ctx_data data = { |
1508 | .hctx = hctx, | |
1509 | .list = list, | |
1510 | }; | |
1429d7c9 | 1511 | |
88459642 | 1512 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1513 | } |
2c3ad667 | 1514 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1515 | |
b347689f ML |
1516 | struct dispatch_rq_data { |
1517 | struct blk_mq_hw_ctx *hctx; | |
1518 | struct request *rq; | |
1519 | }; | |
1520 | ||
1521 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1522 | void *data) | |
1523 | { | |
1524 | struct dispatch_rq_data *dispatch_data = data; | |
1525 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1526 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1527 | enum hctx_type type = hctx->type; |
b347689f ML |
1528 | |
1529 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1530 | if (!list_empty(&ctx->rq_lists[type])) { |
1531 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1532 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1533 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1534 | sbitmap_clear_bit(sb, bitnr); |
1535 | } | |
1536 | spin_unlock(&ctx->lock); | |
1537 | ||
1538 | return !dispatch_data->rq; | |
1539 | } | |
1540 | ||
1541 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1542 | struct blk_mq_ctx *start) | |
1543 | { | |
f31967f0 | 1544 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1545 | struct dispatch_rq_data data = { |
1546 | .hctx = hctx, | |
1547 | .rq = NULL, | |
1548 | }; | |
1549 | ||
1550 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1551 | dispatch_rq_from_ctx, &data); | |
1552 | ||
1553 | return data.rq; | |
1554 | } | |
1555 | ||
a808a9d5 | 1556 | static bool __blk_mq_alloc_driver_tag(struct request *rq) |
570e9b73 | 1557 | { |
ae0f1a73 | 1558 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1559 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1560 | int tag; |
1561 | ||
568f2700 ML |
1562 | blk_mq_tag_busy(rq->mq_hctx); |
1563 | ||
570e9b73 | 1564 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1565 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1566 | tag_offset = 0; |
28500850 ML |
1567 | } else { |
1568 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1569 | return false; | |
570e9b73 ML |
1570 | } |
1571 | ||
570e9b73 ML |
1572 | tag = __sbitmap_queue_get(bt); |
1573 | if (tag == BLK_MQ_NO_TAG) | |
1574 | return false; | |
1575 | ||
1576 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1577 | return true; |
1578 | } | |
1579 | ||
a808a9d5 | 1580 | bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq) |
570e9b73 | 1581 | { |
a808a9d5 | 1582 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq)) |
568f2700 ML |
1583 | return false; |
1584 | ||
51db1c37 | 1585 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1586 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1587 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1588 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1589 | } |
1590 | hctx->tags->rqs[rq->tag] = rq; | |
1591 | return true; | |
570e9b73 ML |
1592 | } |
1593 | ||
eb619fdb JA |
1594 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1595 | int flags, void *key) | |
da55f2cc OS |
1596 | { |
1597 | struct blk_mq_hw_ctx *hctx; | |
1598 | ||
1599 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1600 | ||
5815839b | 1601 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1602 | if (!list_empty(&wait->entry)) { |
1603 | struct sbitmap_queue *sbq; | |
1604 | ||
1605 | list_del_init(&wait->entry); | |
ae0f1a73 | 1606 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1607 | atomic_dec(&sbq->ws_active); |
1608 | } | |
5815839b ML |
1609 | spin_unlock(&hctx->dispatch_wait_lock); |
1610 | ||
da55f2cc OS |
1611 | blk_mq_run_hw_queue(hctx, true); |
1612 | return 1; | |
1613 | } | |
1614 | ||
f906a6a0 JA |
1615 | /* |
1616 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1617 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1618 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1619 | * marking us as waiting. |
1620 | */ | |
2278d69f | 1621 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1622 | struct request *rq) |
da55f2cc | 1623 | { |
ae0f1a73 | 1624 | struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags; |
5815839b | 1625 | struct wait_queue_head *wq; |
f906a6a0 JA |
1626 | wait_queue_entry_t *wait; |
1627 | bool ret; | |
da55f2cc | 1628 | |
51db1c37 | 1629 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
684b7324 | 1630 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1631 | |
c27d53fb BVA |
1632 | /* |
1633 | * It's possible that a tag was freed in the window between the | |
1634 | * allocation failure and adding the hardware queue to the wait | |
1635 | * queue. | |
1636 | * | |
1637 | * Don't clear RESTART here, someone else could have set it. | |
1638 | * At most this will cost an extra queue run. | |
1639 | */ | |
8ab6bb9e | 1640 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1641 | } |
eb619fdb | 1642 | |
2278d69f | 1643 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1644 | if (!list_empty_careful(&wait->entry)) |
1645 | return false; | |
1646 | ||
e8618575 | 1647 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1648 | |
1649 | spin_lock_irq(&wq->lock); | |
1650 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1651 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1652 | spin_unlock(&hctx->dispatch_wait_lock); |
1653 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1654 | return false; |
eb619fdb JA |
1655 | } |
1656 | ||
e8618575 | 1657 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1658 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1659 | __add_wait_queue(wq, wait); | |
c27d53fb | 1660 | |
da55f2cc | 1661 | /* |
eb619fdb JA |
1662 | * It's possible that a tag was freed in the window between the |
1663 | * allocation failure and adding the hardware queue to the wait | |
1664 | * queue. | |
da55f2cc | 1665 | */ |
8ab6bb9e | 1666 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1667 | if (!ret) { |
5815839b ML |
1668 | spin_unlock(&hctx->dispatch_wait_lock); |
1669 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1670 | return false; |
eb619fdb | 1671 | } |
c27d53fb BVA |
1672 | |
1673 | /* | |
1674 | * We got a tag, remove ourselves from the wait queue to ensure | |
1675 | * someone else gets the wakeup. | |
1676 | */ | |
c27d53fb | 1677 | list_del_init(&wait->entry); |
e8618575 | 1678 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1679 | spin_unlock(&hctx->dispatch_wait_lock); |
1680 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1681 | |
1682 | return true; | |
da55f2cc OS |
1683 | } |
1684 | ||
6e768717 ML |
1685 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1686 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1687 | /* | |
1688 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1689 | * - EWMA is one simple way to compute running average value | |
1690 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1691 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1692 | * factor doesn't matter because EWMA decreases exponentially | |
1693 | */ | |
1694 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1695 | { | |
1696 | unsigned int ewma; | |
1697 | ||
6e768717 ML |
1698 | ewma = hctx->dispatch_busy; |
1699 | ||
1700 | if (!ewma && !busy) | |
1701 | return; | |
1702 | ||
1703 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1704 | if (busy) | |
1705 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1706 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1707 | ||
1708 | hctx->dispatch_busy = ewma; | |
1709 | } | |
1710 | ||
86ff7c2a ML |
1711 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1712 | ||
c92a4103 JT |
1713 | static void blk_mq_handle_dev_resource(struct request *rq, |
1714 | struct list_head *list) | |
1715 | { | |
1716 | struct request *next = | |
1717 | list_first_entry_or_null(list, struct request, queuelist); | |
1718 | ||
1719 | /* | |
1720 | * If an I/O scheduler has been configured and we got a driver tag for | |
1721 | * the next request already, free it. | |
1722 | */ | |
1723 | if (next) | |
1724 | blk_mq_put_driver_tag(next); | |
1725 | ||
1726 | list_add(&rq->queuelist, list); | |
1727 | __blk_mq_requeue_request(rq); | |
1728 | } | |
1729 | ||
0512a75b KB |
1730 | static void blk_mq_handle_zone_resource(struct request *rq, |
1731 | struct list_head *zone_list) | |
1732 | { | |
1733 | /* | |
1734 | * If we end up here it is because we cannot dispatch a request to a | |
1735 | * specific zone due to LLD level zone-write locking or other zone | |
1736 | * related resource not being available. In this case, set the request | |
1737 | * aside in zone_list for retrying it later. | |
1738 | */ | |
1739 | list_add(&rq->queuelist, zone_list); | |
1740 | __blk_mq_requeue_request(rq); | |
1741 | } | |
1742 | ||
75383524 ML |
1743 | enum prep_dispatch { |
1744 | PREP_DISPATCH_OK, | |
1745 | PREP_DISPATCH_NO_TAG, | |
1746 | PREP_DISPATCH_NO_BUDGET, | |
1747 | }; | |
1748 | ||
1749 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1750 | bool need_budget) | |
1751 | { | |
1752 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1753 | int budget_token = -1; |
75383524 | 1754 | |
2a5a24aa ML |
1755 | if (need_budget) { |
1756 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1757 | if (budget_token < 0) { | |
1758 | blk_mq_put_driver_tag(rq); | |
1759 | return PREP_DISPATCH_NO_BUDGET; | |
1760 | } | |
1761 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1762 | } |
1763 | ||
1764 | if (!blk_mq_get_driver_tag(rq)) { | |
1765 | /* | |
1766 | * The initial allocation attempt failed, so we need to | |
1767 | * rerun the hardware queue when a tag is freed. The | |
1768 | * waitqueue takes care of that. If the queue is run | |
1769 | * before we add this entry back on the dispatch list, | |
1770 | * we'll re-run it below. | |
1771 | */ | |
1772 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1773 | /* |
1774 | * All budgets not got from this function will be put | |
1775 | * together during handling partial dispatch | |
1776 | */ | |
1777 | if (need_budget) | |
2a5a24aa | 1778 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1779 | return PREP_DISPATCH_NO_TAG; |
1780 | } | |
1781 | } | |
1782 | ||
1783 | return PREP_DISPATCH_OK; | |
1784 | } | |
1785 | ||
1fd40b5e ML |
1786 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1787 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1788 | struct list_head *list) |
1fd40b5e | 1789 | { |
2a5a24aa | 1790 | struct request *rq; |
1fd40b5e | 1791 | |
2a5a24aa ML |
1792 | list_for_each_entry(rq, list, queuelist) { |
1793 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1794 | |
2a5a24aa ML |
1795 | if (budget_token >= 0) |
1796 | blk_mq_put_dispatch_budget(q, budget_token); | |
1797 | } | |
1fd40b5e ML |
1798 | } |
1799 | ||
1f57f8d4 JA |
1800 | /* |
1801 | * Returns true if we did some work AND can potentially do more. | |
1802 | */ | |
445874e8 | 1803 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1804 | unsigned int nr_budgets) |
320ae51f | 1805 | { |
75383524 | 1806 | enum prep_dispatch prep; |
445874e8 | 1807 | struct request_queue *q = hctx->queue; |
6d6f167c | 1808 | struct request *rq, *nxt; |
fc17b653 | 1809 | int errors, queued; |
86ff7c2a | 1810 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 1811 | LIST_HEAD(zone_list); |
9586e67b | 1812 | bool needs_resource = false; |
320ae51f | 1813 | |
81380ca1 OS |
1814 | if (list_empty(list)) |
1815 | return false; | |
1816 | ||
320ae51f JA |
1817 | /* |
1818 | * Now process all the entries, sending them to the driver. | |
1819 | */ | |
93efe981 | 1820 | errors = queued = 0; |
81380ca1 | 1821 | do { |
74c45052 | 1822 | struct blk_mq_queue_data bd; |
320ae51f | 1823 | |
f04c3df3 | 1824 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1825 | |
445874e8 | 1826 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 1827 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 1828 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 1829 | break; |
de148297 | 1830 | |
320ae51f | 1831 | list_del_init(&rq->queuelist); |
320ae51f | 1832 | |
74c45052 | 1833 | bd.rq = rq; |
113285b4 JA |
1834 | |
1835 | /* | |
1836 | * Flag last if we have no more requests, or if we have more | |
1837 | * but can't assign a driver tag to it. | |
1838 | */ | |
1839 | if (list_empty(list)) | |
1840 | bd.last = true; | |
1841 | else { | |
113285b4 | 1842 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1843 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1844 | } |
74c45052 | 1845 | |
1fd40b5e ML |
1846 | /* |
1847 | * once the request is queued to lld, no need to cover the | |
1848 | * budget any more | |
1849 | */ | |
1850 | if (nr_budgets) | |
1851 | nr_budgets--; | |
74c45052 | 1852 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
1853 | switch (ret) { |
1854 | case BLK_STS_OK: | |
1855 | queued++; | |
320ae51f | 1856 | break; |
7bf13729 | 1857 | case BLK_STS_RESOURCE: |
9586e67b NA |
1858 | needs_resource = true; |
1859 | fallthrough; | |
7bf13729 ML |
1860 | case BLK_STS_DEV_RESOURCE: |
1861 | blk_mq_handle_dev_resource(rq, list); | |
1862 | goto out; | |
1863 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
1864 | /* |
1865 | * Move the request to zone_list and keep going through | |
1866 | * the dispatch list to find more requests the drive can | |
1867 | * accept. | |
1868 | */ | |
1869 | blk_mq_handle_zone_resource(rq, &zone_list); | |
9586e67b | 1870 | needs_resource = true; |
7bf13729 ML |
1871 | break; |
1872 | default: | |
93efe981 | 1873 | errors++; |
e21ee5a6 | 1874 | blk_mq_end_request(rq, ret); |
320ae51f | 1875 | } |
81380ca1 | 1876 | } while (!list_empty(list)); |
7bf13729 | 1877 | out: |
0512a75b KB |
1878 | if (!list_empty(&zone_list)) |
1879 | list_splice_tail_init(&zone_list, list); | |
1880 | ||
632bfb63 | 1881 | /* If we didn't flush the entire list, we could have told the driver |
1882 | * there was more coming, but that turned out to be a lie. | |
1883 | */ | |
1884 | if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued) | |
1885 | q->mq_ops->commit_rqs(hctx); | |
320ae51f JA |
1886 | /* |
1887 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1888 | * that is where we will continue on next queue run. | |
1889 | */ | |
f04c3df3 | 1890 | if (!list_empty(list)) { |
86ff7c2a | 1891 | bool needs_restart; |
75383524 ML |
1892 | /* For non-shared tags, the RESTART check will suffice */ |
1893 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
51db1c37 | 1894 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); |
86ff7c2a | 1895 | |
2a5a24aa ML |
1896 | if (nr_budgets) |
1897 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 1898 | |
320ae51f | 1899 | spin_lock(&hctx->lock); |
01e99aec | 1900 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 1901 | spin_unlock(&hctx->lock); |
f04c3df3 | 1902 | |
d7d8535f ML |
1903 | /* |
1904 | * Order adding requests to hctx->dispatch and checking | |
1905 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
1906 | * in blk_mq_sched_restart(). Avoid restart code path to | |
1907 | * miss the new added requests to hctx->dispatch, meantime | |
1908 | * SCHED_RESTART is observed here. | |
1909 | */ | |
1910 | smp_mb(); | |
1911 | ||
9ba52e58 | 1912 | /* |
710c785f BVA |
1913 | * If SCHED_RESTART was set by the caller of this function and |
1914 | * it is no longer set that means that it was cleared by another | |
1915 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1916 | * |
eb619fdb JA |
1917 | * If 'no_tag' is set, that means that we failed getting |
1918 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1919 | * waitqueue is no longer active, ensure that we run the queue | |
1920 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1921 | * |
710c785f BVA |
1922 | * If no I/O scheduler has been configured it is possible that |
1923 | * the hardware queue got stopped and restarted before requests | |
1924 | * were pushed back onto the dispatch list. Rerun the queue to | |
1925 | * avoid starvation. Notes: | |
1926 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1927 | * been stopped before rerunning a queue. | |
1928 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1929 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1930 | * and dm-rq. |
86ff7c2a ML |
1931 | * |
1932 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1933 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 | 1934 | * that could otherwise occur if the queue is idle. We'll do |
9586e67b NA |
1935 | * similar if we couldn't get budget or couldn't lock a zone |
1936 | * and SCHED_RESTART is set. | |
bd166ef1 | 1937 | */ |
86ff7c2a | 1938 | needs_restart = blk_mq_sched_needs_restart(hctx); |
9586e67b NA |
1939 | if (prep == PREP_DISPATCH_NO_BUDGET) |
1940 | needs_resource = true; | |
86ff7c2a | 1941 | if (!needs_restart || |
eb619fdb | 1942 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1943 | blk_mq_run_hw_queue(hctx, true); |
9586e67b | 1944 | else if (needs_restart && needs_resource) |
86ff7c2a | 1945 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 1946 | |
6e768717 | 1947 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1948 | return false; |
6e768717 ML |
1949 | } else |
1950 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1951 | |
93efe981 | 1952 | return (queued + errors) != 0; |
f04c3df3 JA |
1953 | } |
1954 | ||
105663f7 AA |
1955 | /** |
1956 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
1957 | * @hctx: Pointer to the hardware queue to run. | |
1958 | * | |
1959 | * Send pending requests to the hardware. | |
1960 | */ | |
6a83e74d BVA |
1961 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1962 | { | |
1963 | int srcu_idx; | |
1964 | ||
b7a71e66 JA |
1965 | /* |
1966 | * We can't run the queue inline with ints disabled. Ensure that | |
1967 | * we catch bad users of this early. | |
1968 | */ | |
1969 | WARN_ON_ONCE(in_interrupt()); | |
1970 | ||
04ced159 | 1971 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1972 | |
04ced159 JA |
1973 | hctx_lock(hctx, &srcu_idx); |
1974 | blk_mq_sched_dispatch_requests(hctx); | |
1975 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1976 | } |
1977 | ||
f82ddf19 ML |
1978 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1979 | { | |
1980 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1981 | ||
1982 | if (cpu >= nr_cpu_ids) | |
1983 | cpu = cpumask_first(hctx->cpumask); | |
1984 | return cpu; | |
1985 | } | |
1986 | ||
506e931f JA |
1987 | /* |
1988 | * It'd be great if the workqueue API had a way to pass | |
1989 | * in a mask and had some smarts for more clever placement. | |
1990 | * For now we just round-robin here, switching for every | |
1991 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1992 | */ | |
1993 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1994 | { | |
7bed4595 | 1995 | bool tried = false; |
476f8c98 | 1996 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1997 | |
b657d7e6 CH |
1998 | if (hctx->queue->nr_hw_queues == 1) |
1999 | return WORK_CPU_UNBOUND; | |
506e931f JA |
2000 | |
2001 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 2002 | select_cpu: |
476f8c98 | 2003 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 2004 | cpu_online_mask); |
506e931f | 2005 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 2006 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2007 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2008 | } | |
2009 | ||
7bed4595 ML |
2010 | /* |
2011 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
2012 | * and it should only happen in the path of handling CPU DEAD. | |
2013 | */ | |
476f8c98 | 2014 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
2015 | if (!tried) { |
2016 | tried = true; | |
2017 | goto select_cpu; | |
2018 | } | |
2019 | ||
2020 | /* | |
2021 | * Make sure to re-select CPU next time once after CPUs | |
2022 | * in hctx->cpumask become online again. | |
2023 | */ | |
476f8c98 | 2024 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
2025 | hctx->next_cpu_batch = 1; |
2026 | return WORK_CPU_UNBOUND; | |
2027 | } | |
476f8c98 ML |
2028 | |
2029 | hctx->next_cpu = next_cpu; | |
2030 | return next_cpu; | |
506e931f JA |
2031 | } |
2032 | ||
105663f7 AA |
2033 | /** |
2034 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
2035 | * @hctx: Pointer to the hardware queue to run. | |
2036 | * @async: If we want to run the queue asynchronously. | |
fa94ba8a | 2037 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
2038 | * |
2039 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
2040 | * with a delay of @msecs. | |
2041 | */ | |
7587a5ae BVA |
2042 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
2043 | unsigned long msecs) | |
320ae51f | 2044 | { |
5435c023 | 2045 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
2046 | return; |
2047 | ||
1b792f2f | 2048 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
2049 | int cpu = get_cpu(); |
2050 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 2051 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 2052 | put_cpu(); |
398205b8 PB |
2053 | return; |
2054 | } | |
e4043dcf | 2055 | |
2a90d4aa | 2056 | put_cpu(); |
e4043dcf | 2057 | } |
398205b8 | 2058 | |
ae943d20 BVA |
2059 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
2060 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
2061 | } |
2062 | ||
105663f7 AA |
2063 | /** |
2064 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
2065 | * @hctx: Pointer to the hardware queue to run. | |
fa94ba8a | 2066 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
2067 | * |
2068 | * Run a hardware queue asynchronously with a delay of @msecs. | |
2069 | */ | |
7587a5ae BVA |
2070 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
2071 | { | |
2072 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
2073 | } | |
2074 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
2075 | ||
105663f7 AA |
2076 | /** |
2077 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
2078 | * @hctx: Pointer to the hardware queue to run. | |
2079 | * @async: If we want to run the queue asynchronously. | |
2080 | * | |
2081 | * Check if the request queue is not in a quiesced state and if there are | |
2082 | * pending requests to be sent. If this is true, run the queue to send requests | |
2083 | * to hardware. | |
2084 | */ | |
626fb735 | 2085 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 2086 | { |
24f5a90f ML |
2087 | int srcu_idx; |
2088 | bool need_run; | |
2089 | ||
2090 | /* | |
2091 | * When queue is quiesced, we may be switching io scheduler, or | |
2092 | * updating nr_hw_queues, or other things, and we can't run queue | |
2093 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
2094 | * | |
2095 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
2096 | * quiesced. | |
2097 | */ | |
04ced159 JA |
2098 | hctx_lock(hctx, &srcu_idx); |
2099 | need_run = !blk_queue_quiesced(hctx->queue) && | |
2100 | blk_mq_hctx_has_pending(hctx); | |
2101 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f | 2102 | |
626fb735 | 2103 | if (need_run) |
79f720a7 | 2104 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 2105 | } |
5b727272 | 2106 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 2107 | |
b6e68ee8 JK |
2108 | /* |
2109 | * Is the request queue handled by an IO scheduler that does not respect | |
2110 | * hardware queues when dispatching? | |
2111 | */ | |
2112 | static bool blk_mq_has_sqsched(struct request_queue *q) | |
2113 | { | |
2114 | struct elevator_queue *e = q->elevator; | |
2115 | ||
2116 | if (e && e->type->ops.dispatch_request && | |
2117 | !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE)) | |
2118 | return true; | |
2119 | return false; | |
2120 | } | |
2121 | ||
2122 | /* | |
2123 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
2124 | * scheduler. | |
2125 | */ | |
2126 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
2127 | { | |
2128 | struct blk_mq_hw_ctx *hctx; | |
2129 | ||
2130 | /* | |
2131 | * If the IO scheduler does not respect hardware queues when | |
2132 | * dispatching, we just don't bother with multiple HW queues and | |
2133 | * dispatch from hctx for the current CPU since running multiple queues | |
2134 | * just causes lock contention inside the scheduler and pointless cache | |
2135 | * bouncing. | |
2136 | */ | |
2137 | hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT, | |
2138 | raw_smp_processor_id()); | |
2139 | if (!blk_mq_hctx_stopped(hctx)) | |
2140 | return hctx; | |
2141 | return NULL; | |
2142 | } | |
2143 | ||
105663f7 | 2144 | /** |
24f7bb88 | 2145 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
2146 | * @q: Pointer to the request queue to run. |
2147 | * @async: If we want to run the queue asynchronously. | |
2148 | */ | |
b94ec296 | 2149 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 2150 | { |
b6e68ee8 | 2151 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
320ae51f JA |
2152 | int i; |
2153 | ||
b6e68ee8 JK |
2154 | sq_hctx = NULL; |
2155 | if (blk_mq_has_sqsched(q)) | |
2156 | sq_hctx = blk_mq_get_sq_hctx(q); | |
320ae51f | 2157 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 2158 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 2159 | continue; |
b6e68ee8 JK |
2160 | /* |
2161 | * Dispatch from this hctx either if there's no hctx preferred | |
2162 | * by IO scheduler or if it has requests that bypass the | |
2163 | * scheduler. | |
2164 | */ | |
2165 | if (!sq_hctx || sq_hctx == hctx || | |
2166 | !list_empty_careful(&hctx->dispatch)) | |
2167 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
2168 | } |
2169 | } | |
b94ec296 | 2170 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 2171 | |
b9151e7b DA |
2172 | /** |
2173 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
2174 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 2175 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
2176 | */ |
2177 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
2178 | { | |
b6e68ee8 | 2179 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
b9151e7b DA |
2180 | int i; |
2181 | ||
b6e68ee8 JK |
2182 | sq_hctx = NULL; |
2183 | if (blk_mq_has_sqsched(q)) | |
2184 | sq_hctx = blk_mq_get_sq_hctx(q); | |
b9151e7b DA |
2185 | queue_for_each_hw_ctx(q, hctx, i) { |
2186 | if (blk_mq_hctx_stopped(hctx)) | |
2187 | continue; | |
b6e68ee8 JK |
2188 | /* |
2189 | * Dispatch from this hctx either if there's no hctx preferred | |
2190 | * by IO scheduler or if it has requests that bypass the | |
2191 | * scheduler. | |
2192 | */ | |
2193 | if (!sq_hctx || sq_hctx == hctx || | |
2194 | !list_empty_careful(&hctx->dispatch)) | |
2195 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
2196 | } |
2197 | } | |
2198 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
2199 | ||
fd001443 BVA |
2200 | /** |
2201 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
2202 | * @q: request queue. | |
2203 | * | |
2204 | * The caller is responsible for serializing this function against | |
2205 | * blk_mq_{start,stop}_hw_queue(). | |
2206 | */ | |
2207 | bool blk_mq_queue_stopped(struct request_queue *q) | |
2208 | { | |
2209 | struct blk_mq_hw_ctx *hctx; | |
2210 | int i; | |
2211 | ||
2212 | queue_for_each_hw_ctx(q, hctx, i) | |
2213 | if (blk_mq_hctx_stopped(hctx)) | |
2214 | return true; | |
2215 | ||
2216 | return false; | |
2217 | } | |
2218 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
2219 | ||
39a70c76 ML |
2220 | /* |
2221 | * This function is often used for pausing .queue_rq() by driver when | |
2222 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2223 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2224 | * |
2225 | * We do not guarantee that dispatch can be drained or blocked | |
2226 | * after blk_mq_stop_hw_queue() returns. Please use | |
2227 | * blk_mq_quiesce_queue() for that requirement. | |
2228 | */ | |
2719aa21 JA |
2229 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
2230 | { | |
641a9ed6 | 2231 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 2232 | |
641a9ed6 | 2233 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 2234 | } |
641a9ed6 | 2235 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 2236 | |
39a70c76 ML |
2237 | /* |
2238 | * This function is often used for pausing .queue_rq() by driver when | |
2239 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 2240 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
2241 | * |
2242 | * We do not guarantee that dispatch can be drained or blocked | |
2243 | * after blk_mq_stop_hw_queues() returns. Please use | |
2244 | * blk_mq_quiesce_queue() for that requirement. | |
2245 | */ | |
2719aa21 JA |
2246 | void blk_mq_stop_hw_queues(struct request_queue *q) |
2247 | { | |
641a9ed6 ML |
2248 | struct blk_mq_hw_ctx *hctx; |
2249 | int i; | |
2250 | ||
2251 | queue_for_each_hw_ctx(q, hctx, i) | |
2252 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
2253 | } |
2254 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
2255 | ||
320ae51f JA |
2256 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
2257 | { | |
2258 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 2259 | |
0ffbce80 | 2260 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
2261 | } |
2262 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
2263 | ||
2f268556 CH |
2264 | void blk_mq_start_hw_queues(struct request_queue *q) |
2265 | { | |
2266 | struct blk_mq_hw_ctx *hctx; | |
2267 | int i; | |
2268 | ||
2269 | queue_for_each_hw_ctx(q, hctx, i) | |
2270 | blk_mq_start_hw_queue(hctx); | |
2271 | } | |
2272 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
2273 | ||
ae911c5e JA |
2274 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
2275 | { | |
2276 | if (!blk_mq_hctx_stopped(hctx)) | |
2277 | return; | |
2278 | ||
2279 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
2280 | blk_mq_run_hw_queue(hctx, async); | |
2281 | } | |
2282 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
2283 | ||
1b4a3258 | 2284 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
2285 | { |
2286 | struct blk_mq_hw_ctx *hctx; | |
2287 | int i; | |
2288 | ||
ae911c5e JA |
2289 | queue_for_each_hw_ctx(q, hctx, i) |
2290 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
2291 | } |
2292 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
2293 | ||
70f4db63 | 2294 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
2295 | { |
2296 | struct blk_mq_hw_ctx *hctx; | |
2297 | ||
9f993737 | 2298 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 2299 | |
21c6e939 | 2300 | /* |
15fe8a90 | 2301 | * If we are stopped, don't run the queue. |
21c6e939 | 2302 | */ |
0841031a | 2303 | if (blk_mq_hctx_stopped(hctx)) |
0196d6b4 | 2304 | return; |
7587a5ae BVA |
2305 | |
2306 | __blk_mq_run_hw_queue(hctx); | |
2307 | } | |
2308 | ||
cfd0c552 | 2309 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
2310 | struct request *rq, |
2311 | bool at_head) | |
320ae51f | 2312 | { |
e57690fe | 2313 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 2314 | enum hctx_type type = hctx->type; |
e57690fe | 2315 | |
7b607814 BVA |
2316 | lockdep_assert_held(&ctx->lock); |
2317 | ||
a54895fa | 2318 | trace_block_rq_insert(rq); |
01b983c9 | 2319 | |
72a0a36e | 2320 | if (at_head) |
c16d6b5a | 2321 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 2322 | else |
c16d6b5a | 2323 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 2324 | } |
4bb659b1 | 2325 | |
2c3ad667 JA |
2326 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
2327 | bool at_head) | |
cfd0c552 ML |
2328 | { |
2329 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
2330 | ||
7b607814 BVA |
2331 | lockdep_assert_held(&ctx->lock); |
2332 | ||
e57690fe | 2333 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 2334 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
2335 | } |
2336 | ||
105663f7 AA |
2337 | /** |
2338 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
2339 | * @rq: Pointer to request to be inserted. | |
26bfeb26 | 2340 | * @at_head: true if the request should be inserted at the head of the list. |
105663f7 AA |
2341 | * @run_queue: If we should run the hardware queue after inserting the request. |
2342 | * | |
157f377b JA |
2343 | * Should only be used carefully, when the caller knows we want to |
2344 | * bypass a potential IO scheduler on the target device. | |
2345 | */ | |
01e99aec ML |
2346 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
2347 | bool run_queue) | |
157f377b | 2348 | { |
ea4f995e | 2349 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
2350 | |
2351 | spin_lock(&hctx->lock); | |
01e99aec ML |
2352 | if (at_head) |
2353 | list_add(&rq->queuelist, &hctx->dispatch); | |
2354 | else | |
2355 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
2356 | spin_unlock(&hctx->lock); |
2357 | ||
b0850297 ML |
2358 | if (run_queue) |
2359 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
2360 | } |
2361 | ||
bd166ef1 JA |
2362 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
2363 | struct list_head *list) | |
320ae51f JA |
2364 | |
2365 | { | |
3f0cedc7 | 2366 | struct request *rq; |
c16d6b5a | 2367 | enum hctx_type type = hctx->type; |
3f0cedc7 | 2368 | |
320ae51f JA |
2369 | /* |
2370 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
2371 | * offline now | |
2372 | */ | |
3f0cedc7 | 2373 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 2374 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 2375 | trace_block_rq_insert(rq); |
320ae51f | 2376 | } |
3f0cedc7 ML |
2377 | |
2378 | spin_lock(&ctx->lock); | |
c16d6b5a | 2379 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 2380 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 2381 | spin_unlock(&ctx->lock); |
320ae51f JA |
2382 | } |
2383 | ||
dc5fc361 JA |
2384 | static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued, |
2385 | bool from_schedule) | |
320ae51f | 2386 | { |
dc5fc361 JA |
2387 | if (hctx->queue->mq_ops->commit_rqs) { |
2388 | trace_block_unplug(hctx->queue, *queued, !from_schedule); | |
2389 | hctx->queue->mq_ops->commit_rqs(hctx); | |
2390 | } | |
2391 | *queued = 0; | |
2392 | } | |
320ae51f | 2393 | |
14ccb66b CH |
2394 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
2395 | unsigned int nr_segs) | |
320ae51f | 2396 | { |
93f221ae EB |
2397 | int err; |
2398 | ||
f924cdde CH |
2399 | if (bio->bi_opf & REQ_RAHEAD) |
2400 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
2401 | ||
2402 | rq->__sector = bio->bi_iter.bi_sector; | |
2403 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 2404 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
2405 | |
2406 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2407 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2408 | WARN_ON_ONCE(err); | |
4b570521 | 2409 | |
b5af37ab | 2410 | blk_account_io_start(rq); |
320ae51f JA |
2411 | } |
2412 | ||
0f95549c | 2413 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2414 | struct request *rq, bool last) |
f984df1f | 2415 | { |
f984df1f | 2416 | struct request_queue *q = rq->q; |
f984df1f SL |
2417 | struct blk_mq_queue_data bd = { |
2418 | .rq = rq, | |
be94f058 | 2419 | .last = last, |
f984df1f | 2420 | }; |
f06345ad | 2421 | blk_status_t ret; |
0f95549c | 2422 | |
0f95549c MS |
2423 | /* |
2424 | * For OK queue, we are done. For error, caller may kill it. | |
2425 | * Any other error (busy), just add it to our list as we | |
2426 | * previously would have done. | |
2427 | */ | |
2428 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2429 | switch (ret) { | |
2430 | case BLK_STS_OK: | |
6ce3dd6e | 2431 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2432 | break; |
2433 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2434 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2435 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2436 | __blk_mq_requeue_request(rq); |
2437 | break; | |
2438 | default: | |
6ce3dd6e | 2439 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2440 | break; |
2441 | } | |
2442 | ||
2443 | return ret; | |
2444 | } | |
2445 | ||
fd9c40f6 | 2446 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 2447 | struct request *rq, |
fd9c40f6 | 2448 | bool bypass_insert, bool last) |
0f95549c MS |
2449 | { |
2450 | struct request_queue *q = rq->q; | |
d964f04a | 2451 | bool run_queue = true; |
2a5a24aa | 2452 | int budget_token; |
d964f04a | 2453 | |
23d4ee19 | 2454 | /* |
fd9c40f6 | 2455 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 2456 | * |
fd9c40f6 BVA |
2457 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
2458 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2459 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 2460 | */ |
fd9c40f6 | 2461 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 2462 | run_queue = false; |
fd9c40f6 BVA |
2463 | bypass_insert = false; |
2464 | goto insert; | |
d964f04a | 2465 | } |
f984df1f | 2466 | |
2ff0682d | 2467 | if ((rq->rq_flags & RQF_ELV) && !bypass_insert) |
fd9c40f6 | 2468 | goto insert; |
2253efc8 | 2469 | |
2a5a24aa ML |
2470 | budget_token = blk_mq_get_dispatch_budget(q); |
2471 | if (budget_token < 0) | |
fd9c40f6 | 2472 | goto insert; |
bd166ef1 | 2473 | |
2a5a24aa ML |
2474 | blk_mq_set_rq_budget_token(rq, budget_token); |
2475 | ||
8ab6bb9e | 2476 | if (!blk_mq_get_driver_tag(rq)) { |
2a5a24aa | 2477 | blk_mq_put_dispatch_budget(q, budget_token); |
fd9c40f6 | 2478 | goto insert; |
88022d72 | 2479 | } |
de148297 | 2480 | |
3e08773c | 2481 | return __blk_mq_issue_directly(hctx, rq, last); |
fd9c40f6 BVA |
2482 | insert: |
2483 | if (bypass_insert) | |
2484 | return BLK_STS_RESOURCE; | |
2485 | ||
db03f88f ML |
2486 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2487 | ||
fd9c40f6 BVA |
2488 | return BLK_STS_OK; |
2489 | } | |
2490 | ||
105663f7 AA |
2491 | /** |
2492 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2493 | * @hctx: Pointer of the associated hardware queue. | |
2494 | * @rq: Pointer to request to be sent. | |
105663f7 AA |
2495 | * |
2496 | * If the device has enough resources to accept a new request now, send the | |
2497 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2498 | * we can try send it another time in the future. Requests inserted at this | |
2499 | * queue have higher priority. | |
2500 | */ | |
fd9c40f6 | 2501 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2502 | struct request *rq) |
fd9c40f6 BVA |
2503 | { |
2504 | blk_status_t ret; | |
2505 | int srcu_idx; | |
2506 | ||
2507 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
2508 | ||
2509 | hctx_lock(hctx, &srcu_idx); | |
2510 | ||
3e08773c | 2511 | ret = __blk_mq_try_issue_directly(hctx, rq, false, true); |
fd9c40f6 | 2512 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
01e99aec | 2513 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2514 | else if (ret != BLK_STS_OK) |
2515 | blk_mq_end_request(rq, ret); | |
2516 | ||
2517 | hctx_unlock(hctx, srcu_idx); | |
2518 | } | |
2519 | ||
06c8c691 | 2520 | static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) |
fd9c40f6 BVA |
2521 | { |
2522 | blk_status_t ret; | |
2523 | int srcu_idx; | |
fd9c40f6 BVA |
2524 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
2525 | ||
2526 | hctx_lock(hctx, &srcu_idx); | |
3e08773c | 2527 | ret = __blk_mq_try_issue_directly(hctx, rq, true, last); |
04ced159 | 2528 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
2529 | |
2530 | return ret; | |
5eb6126e CH |
2531 | } |
2532 | ||
b84c5b50 CH |
2533 | static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule) |
2534 | { | |
2535 | struct blk_mq_hw_ctx *hctx = NULL; | |
2536 | struct request *rq; | |
2537 | int queued = 0; | |
2538 | int errors = 0; | |
2539 | ||
2540 | while ((rq = rq_list_pop(&plug->mq_list))) { | |
2541 | bool last = rq_list_empty(plug->mq_list); | |
2542 | blk_status_t ret; | |
2543 | ||
2544 | if (hctx != rq->mq_hctx) { | |
2545 | if (hctx) | |
2546 | blk_mq_commit_rqs(hctx, &queued, from_schedule); | |
2547 | hctx = rq->mq_hctx; | |
2548 | } | |
2549 | ||
2550 | ret = blk_mq_request_issue_directly(rq, last); | |
2551 | switch (ret) { | |
2552 | case BLK_STS_OK: | |
2553 | queued++; | |
2554 | break; | |
2555 | case BLK_STS_RESOURCE: | |
2556 | case BLK_STS_DEV_RESOURCE: | |
2557 | blk_mq_request_bypass_insert(rq, false, last); | |
2558 | blk_mq_commit_rqs(hctx, &queued, from_schedule); | |
2559 | return; | |
2560 | default: | |
2561 | blk_mq_end_request(rq, ret); | |
2562 | errors++; | |
2563 | break; | |
2564 | } | |
2565 | } | |
2566 | ||
2567 | /* | |
2568 | * If we didn't flush the entire list, we could have told the driver | |
2569 | * there was more coming, but that turned out to be a lie. | |
2570 | */ | |
2571 | if (errors) | |
2572 | blk_mq_commit_rqs(hctx, &queued, from_schedule); | |
2573 | } | |
2574 | ||
2575 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
2576 | { | |
2577 | struct blk_mq_hw_ctx *this_hctx; | |
2578 | struct blk_mq_ctx *this_ctx; | |
2579 | unsigned int depth; | |
2580 | LIST_HEAD(list); | |
2581 | ||
2582 | if (rq_list_empty(plug->mq_list)) | |
2583 | return; | |
2584 | plug->rq_count = 0; | |
2585 | ||
2586 | if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) { | |
2587 | blk_mq_plug_issue_direct(plug, false); | |
2588 | if (rq_list_empty(plug->mq_list)) | |
2589 | return; | |
2590 | } | |
2591 | ||
2592 | this_hctx = NULL; | |
2593 | this_ctx = NULL; | |
2594 | depth = 0; | |
2595 | do { | |
2596 | struct request *rq; | |
2597 | ||
2598 | rq = rq_list_pop(&plug->mq_list); | |
2599 | ||
2600 | if (!this_hctx) { | |
2601 | this_hctx = rq->mq_hctx; | |
2602 | this_ctx = rq->mq_ctx; | |
2603 | } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) { | |
2604 | trace_block_unplug(this_hctx->queue, depth, | |
2605 | !from_schedule); | |
2606 | blk_mq_sched_insert_requests(this_hctx, this_ctx, | |
2607 | &list, from_schedule); | |
2608 | depth = 0; | |
2609 | this_hctx = rq->mq_hctx; | |
2610 | this_ctx = rq->mq_ctx; | |
2611 | ||
2612 | } | |
2613 | ||
2614 | list_add(&rq->queuelist, &list); | |
2615 | depth++; | |
2616 | } while (!rq_list_empty(plug->mq_list)); | |
2617 | ||
2618 | if (!list_empty(&list)) { | |
2619 | trace_block_unplug(this_hctx->queue, depth, !from_schedule); | |
2620 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, | |
2621 | from_schedule); | |
2622 | } | |
2623 | } | |
2624 | ||
6ce3dd6e ML |
2625 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2626 | struct list_head *list) | |
2627 | { | |
536167d4 | 2628 | int queued = 0; |
632bfb63 | 2629 | int errors = 0; |
536167d4 | 2630 | |
6ce3dd6e | 2631 | while (!list_empty(list)) { |
fd9c40f6 | 2632 | blk_status_t ret; |
6ce3dd6e ML |
2633 | struct request *rq = list_first_entry(list, struct request, |
2634 | queuelist); | |
2635 | ||
2636 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
2637 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
2638 | if (ret != BLK_STS_OK) { | |
2639 | if (ret == BLK_STS_RESOURCE || | |
2640 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 2641 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 2642 | list_empty(list)); |
fd9c40f6 BVA |
2643 | break; |
2644 | } | |
2645 | blk_mq_end_request(rq, ret); | |
632bfb63 | 2646 | errors++; |
536167d4 KB |
2647 | } else |
2648 | queued++; | |
6ce3dd6e | 2649 | } |
d666ba98 JA |
2650 | |
2651 | /* | |
2652 | * If we didn't flush the entire list, we could have told | |
2653 | * the driver there was more coming, but that turned out to | |
2654 | * be a lie. | |
2655 | */ | |
632bfb63 | 2656 | if ((!list_empty(list) || errors) && |
2657 | hctx->queue->mq_ops->commit_rqs && queued) | |
d666ba98 | 2658 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
2659 | } |
2660 | ||
7f2a6a69 | 2661 | /* |
ba0ffdd8 | 2662 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple |
7f2a6a69 SL |
2663 | * queues. This is important for md arrays to benefit from merging |
2664 | * requests. | |
2665 | */ | |
2666 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
2667 | { | |
2668 | if (plug->multiple_queues) | |
ba0ffdd8 | 2669 | return BLK_MAX_REQUEST_COUNT * 2; |
7f2a6a69 SL |
2670 | return BLK_MAX_REQUEST_COUNT; |
2671 | } | |
2672 | ||
1e9c2303 CH |
2673 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
2674 | { | |
2675 | struct request *last = rq_list_peek(&plug->mq_list); | |
2676 | ||
2677 | if (!plug->rq_count) { | |
2678 | trace_block_plug(rq->q); | |
2679 | } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || | |
2680 | (!blk_queue_nomerges(rq->q) && | |
2681 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
2682 | blk_mq_flush_plug_list(plug, false); | |
2683 | trace_block_plug(rq->q); | |
2684 | } | |
2685 | ||
2686 | if (!plug->multiple_queues && last && last->q != rq->q) | |
2687 | plug->multiple_queues = true; | |
2688 | if (!plug->has_elevator && (rq->rq_flags & RQF_ELV)) | |
2689 | plug->has_elevator = true; | |
2690 | rq->rq_next = NULL; | |
2691 | rq_list_add(&plug->mq_list, rq); | |
2692 | plug->rq_count++; | |
2693 | } | |
2694 | ||
b131f201 | 2695 | static bool blk_mq_attempt_bio_merge(struct request_queue *q, |
0c5bcc92 | 2696 | struct bio *bio, unsigned int nr_segs) |
900e0807 JA |
2697 | { |
2698 | if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { | |
0c5bcc92 | 2699 | if (blk_attempt_plug_merge(q, bio, nr_segs)) |
900e0807 JA |
2700 | return true; |
2701 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) | |
2702 | return true; | |
2703 | } | |
2704 | return false; | |
2705 | } | |
2706 | ||
71539717 JA |
2707 | static struct request *blk_mq_get_new_requests(struct request_queue *q, |
2708 | struct blk_plug *plug, | |
900e0807 | 2709 | struct bio *bio, |
0c5bcc92 | 2710 | unsigned int nsegs) |
71539717 JA |
2711 | { |
2712 | struct blk_mq_alloc_data data = { | |
2713 | .q = q, | |
2714 | .nr_tags = 1, | |
2715 | .cmd_flags = bio->bi_opf, | |
2716 | }; | |
2717 | struct request *rq; | |
2718 | ||
0c5bcc92 | 2719 | if (blk_mq_attempt_bio_merge(q, bio, nsegs)) |
b637108a | 2720 | return NULL; |
900e0807 JA |
2721 | |
2722 | rq_qos_throttle(q, bio); | |
2723 | ||
71539717 JA |
2724 | if (plug) { |
2725 | data.nr_tags = plug->nr_ios; | |
2726 | plug->nr_ios = 1; | |
2727 | data.cached_rq = &plug->cached_rq; | |
2728 | } | |
2729 | ||
2730 | rq = __blk_mq_alloc_requests(&data); | |
2731 | if (rq) | |
2732 | return rq; | |
2733 | ||
2734 | rq_qos_cleanup(q, bio); | |
2735 | if (bio->bi_opf & REQ_NOWAIT) | |
2736 | bio_wouldblock_error(bio); | |
b637108a | 2737 | |
71539717 JA |
2738 | return NULL; |
2739 | } | |
2740 | ||
95febeb6 | 2741 | static inline bool blk_mq_can_use_cached_rq(struct request *rq, struct bio *bio) |
b637108a ML |
2742 | { |
2743 | if (blk_mq_get_hctx_type(bio->bi_opf) != rq->mq_hctx->type) | |
2744 | return false; | |
2745 | ||
2746 | if (op_is_flush(rq->cmd_flags) != op_is_flush(bio->bi_opf)) | |
2747 | return false; | |
2748 | ||
2749 | return true; | |
2750 | } | |
2751 | ||
71539717 JA |
2752 | static inline struct request *blk_mq_get_request(struct request_queue *q, |
2753 | struct blk_plug *plug, | |
900e0807 | 2754 | struct bio *bio, |
0c5bcc92 | 2755 | unsigned int nsegs) |
71539717 | 2756 | { |
b637108a ML |
2757 | struct request *rq; |
2758 | bool checked = false; | |
2759 | ||
71539717 | 2760 | if (plug) { |
71539717 | 2761 | rq = rq_list_peek(&plug->cached_rq); |
10c47870 | 2762 | if (rq && rq->q == q) { |
900e0807 JA |
2763 | if (unlikely(!submit_bio_checks(bio))) |
2764 | return NULL; | |
0c5bcc92 | 2765 | if (blk_mq_attempt_bio_merge(q, bio, nsegs)) |
900e0807 | 2766 | return NULL; |
b637108a ML |
2767 | checked = true; |
2768 | if (!blk_mq_can_use_cached_rq(rq, bio)) | |
2769 | goto fallback; | |
2770 | rq->cmd_flags = bio->bi_opf; | |
71539717 JA |
2771 | plug->cached_rq = rq_list_next(rq); |
2772 | INIT_LIST_HEAD(&rq->queuelist); | |
900e0807 | 2773 | rq_qos_throttle(q, bio); |
71539717 JA |
2774 | return rq; |
2775 | } | |
2776 | } | |
2777 | ||
b637108a ML |
2778 | fallback: |
2779 | if (unlikely(bio_queue_enter(bio))) | |
2780 | return NULL; | |
95febeb6 JA |
2781 | if (unlikely(!checked && !submit_bio_checks(bio))) |
2782 | goto out_put; | |
0c5bcc92 | 2783 | rq = blk_mq_get_new_requests(q, plug, bio, nsegs); |
95febeb6 JA |
2784 | if (rq) |
2785 | return rq; | |
2786 | out_put: | |
2787 | blk_queue_exit(q); | |
2788 | return NULL; | |
71539717 JA |
2789 | } |
2790 | ||
105663f7 | 2791 | /** |
c62b37d9 | 2792 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2793 | * @bio: Bio pointer. |
2794 | * | |
2795 | * Builds up a request structure from @q and @bio and send to the device. The | |
2796 | * request may not be queued directly to hardware if: | |
2797 | * * This request can be merged with another one | |
2798 | * * We want to place request at plug queue for possible future merging | |
2799 | * * There is an IO scheduler active at this queue | |
2800 | * | |
2801 | * It will not queue the request if there is an error with the bio, or at the | |
2802 | * request creation. | |
105663f7 | 2803 | */ |
3e08773c | 2804 | void blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2805 | { |
ed6cddef | 2806 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
ef295ecf | 2807 | const int is_sync = op_is_sync(bio->bi_opf); |
07068d5b | 2808 | struct request *rq; |
f984df1f | 2809 | struct blk_plug *plug; |
abd45c15 | 2810 | unsigned int nr_segs = 1; |
a892c8d5 | 2811 | blk_status_t ret; |
07068d5b | 2812 | |
900e0807 JA |
2813 | if (unlikely(!blk_crypto_bio_prep(&bio))) |
2814 | return; | |
2815 | ||
07068d5b | 2816 | blk_queue_bounce(q, &bio); |
abd45c15 JA |
2817 | if (blk_may_split(q, bio)) |
2818 | __blk_queue_split(q, &bio, &nr_segs); | |
f36ea50c | 2819 | |
e23947bd | 2820 | if (!bio_integrity_prep(bio)) |
900e0807 | 2821 | return; |
87760e5e | 2822 | |
47c122e3 | 2823 | plug = blk_mq_plug(q, bio); |
0c5bcc92 | 2824 | rq = blk_mq_get_request(q, plug, bio, nr_segs); |
71539717 | 2825 | if (unlikely(!rq)) |
900e0807 | 2826 | return; |
87760e5e | 2827 | |
e8a676d6 | 2828 | trace_block_getrq(bio); |
d6f1dda2 | 2829 | |
c1c80384 | 2830 | rq_qos_track(q, rq, bio); |
07068d5b | 2831 | |
970d168d BVA |
2832 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2833 | ||
a892c8d5 ST |
2834 | ret = blk_crypto_init_request(rq); |
2835 | if (ret != BLK_STS_OK) { | |
2836 | bio->bi_status = ret; | |
2837 | bio_endio(bio); | |
2838 | blk_mq_free_request(rq); | |
3e08773c | 2839 | return; |
a892c8d5 ST |
2840 | } |
2841 | ||
2b504bd4 ML |
2842 | if (op_is_flush(bio->bi_opf)) { |
2843 | blk_insert_flush(rq); | |
d92ca9d8 | 2844 | return; |
2b504bd4 | 2845 | } |
d92ca9d8 | 2846 | |
1e9c2303 | 2847 | if (plug) |
ce5b009c | 2848 | blk_add_rq_to_plug(plug, rq); |
1e9c2303 CH |
2849 | else if ((rq->rq_flags & RQF_ELV) || |
2850 | (rq->mq_hctx->dispatch_busy && | |
2851 | (q->nr_hw_queues == 1 || !is_sync))) | |
a12de1d4 | 2852 | blk_mq_sched_insert_request(rq, false, true, true); |
1e9c2303 | 2853 | else |
0c5bcc92 | 2854 | blk_mq_try_issue_directly(rq->mq_hctx, rq); |
320ae51f JA |
2855 | } |
2856 | ||
06c8c691 CH |
2857 | /** |
2858 | * blk_cloned_rq_check_limits - Helper function to check a cloned request | |
2859 | * for the new queue limits | |
2860 | * @q: the queue | |
2861 | * @rq: the request being checked | |
2862 | * | |
2863 | * Description: | |
2864 | * @rq may have been made based on weaker limitations of upper-level queues | |
2865 | * in request stacking drivers, and it may violate the limitation of @q. | |
2866 | * Since the block layer and the underlying device driver trust @rq | |
2867 | * after it is inserted to @q, it should be checked against @q before | |
2868 | * the insertion using this generic function. | |
2869 | * | |
2870 | * Request stacking drivers like request-based dm may change the queue | |
2871 | * limits when retrying requests on other queues. Those requests need | |
2872 | * to be checked against the new queue limits again during dispatch. | |
2873 | */ | |
2874 | static blk_status_t blk_cloned_rq_check_limits(struct request_queue *q, | |
2875 | struct request *rq) | |
2876 | { | |
2877 | unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); | |
2878 | ||
2879 | if (blk_rq_sectors(rq) > max_sectors) { | |
2880 | /* | |
2881 | * SCSI device does not have a good way to return if | |
2882 | * Write Same/Zero is actually supported. If a device rejects | |
2883 | * a non-read/write command (discard, write same,etc.) the | |
2884 | * low-level device driver will set the relevant queue limit to | |
2885 | * 0 to prevent blk-lib from issuing more of the offending | |
2886 | * operations. Commands queued prior to the queue limit being | |
2887 | * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O | |
2888 | * errors being propagated to upper layers. | |
2889 | */ | |
2890 | if (max_sectors == 0) | |
2891 | return BLK_STS_NOTSUPP; | |
2892 | ||
2893 | printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", | |
2894 | __func__, blk_rq_sectors(rq), max_sectors); | |
2895 | return BLK_STS_IOERR; | |
2896 | } | |
2897 | ||
2898 | /* | |
2899 | * The queue settings related to segment counting may differ from the | |
2900 | * original queue. | |
2901 | */ | |
2902 | rq->nr_phys_segments = blk_recalc_rq_segments(rq); | |
2903 | if (rq->nr_phys_segments > queue_max_segments(q)) { | |
2904 | printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n", | |
2905 | __func__, rq->nr_phys_segments, queue_max_segments(q)); | |
2906 | return BLK_STS_IOERR; | |
2907 | } | |
2908 | ||
2909 | return BLK_STS_OK; | |
2910 | } | |
2911 | ||
2912 | /** | |
2913 | * blk_insert_cloned_request - Helper for stacking drivers to submit a request | |
2914 | * @q: the queue to submit the request | |
2915 | * @rq: the request being queued | |
2916 | */ | |
2917 | blk_status_t blk_insert_cloned_request(struct request_queue *q, struct request *rq) | |
2918 | { | |
2919 | blk_status_t ret; | |
2920 | ||
2921 | ret = blk_cloned_rq_check_limits(q, rq); | |
2922 | if (ret != BLK_STS_OK) | |
2923 | return ret; | |
2924 | ||
2925 | if (rq->rq_disk && | |
2926 | should_fail_request(rq->rq_disk->part0, blk_rq_bytes(rq))) | |
2927 | return BLK_STS_IOERR; | |
2928 | ||
2929 | if (blk_crypto_insert_cloned_request(rq)) | |
2930 | return BLK_STS_IOERR; | |
2931 | ||
2932 | blk_account_io_start(rq); | |
2933 | ||
2934 | /* | |
2935 | * Since we have a scheduler attached on the top device, | |
2936 | * bypass a potential scheduler on the bottom device for | |
2937 | * insert. | |
2938 | */ | |
2939 | return blk_mq_request_issue_directly(rq, true); | |
2940 | } | |
2941 | EXPORT_SYMBOL_GPL(blk_insert_cloned_request); | |
2942 | ||
2943 | /** | |
2944 | * blk_rq_unprep_clone - Helper function to free all bios in a cloned request | |
2945 | * @rq: the clone request to be cleaned up | |
2946 | * | |
2947 | * Description: | |
2948 | * Free all bios in @rq for a cloned request. | |
2949 | */ | |
2950 | void blk_rq_unprep_clone(struct request *rq) | |
2951 | { | |
2952 | struct bio *bio; | |
2953 | ||
2954 | while ((bio = rq->bio) != NULL) { | |
2955 | rq->bio = bio->bi_next; | |
2956 | ||
2957 | bio_put(bio); | |
2958 | } | |
2959 | } | |
2960 | EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); | |
2961 | ||
2962 | /** | |
2963 | * blk_rq_prep_clone - Helper function to setup clone request | |
2964 | * @rq: the request to be setup | |
2965 | * @rq_src: original request to be cloned | |
2966 | * @bs: bio_set that bios for clone are allocated from | |
2967 | * @gfp_mask: memory allocation mask for bio | |
2968 | * @bio_ctr: setup function to be called for each clone bio. | |
2969 | * Returns %0 for success, non %0 for failure. | |
2970 | * @data: private data to be passed to @bio_ctr | |
2971 | * | |
2972 | * Description: | |
2973 | * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. | |
2974 | * Also, pages which the original bios are pointing to are not copied | |
2975 | * and the cloned bios just point same pages. | |
2976 | * So cloned bios must be completed before original bios, which means | |
2977 | * the caller must complete @rq before @rq_src. | |
2978 | */ | |
2979 | int blk_rq_prep_clone(struct request *rq, struct request *rq_src, | |
2980 | struct bio_set *bs, gfp_t gfp_mask, | |
2981 | int (*bio_ctr)(struct bio *, struct bio *, void *), | |
2982 | void *data) | |
2983 | { | |
2984 | struct bio *bio, *bio_src; | |
2985 | ||
2986 | if (!bs) | |
2987 | bs = &fs_bio_set; | |
2988 | ||
2989 | __rq_for_each_bio(bio_src, rq_src) { | |
2990 | bio = bio_clone_fast(bio_src, gfp_mask, bs); | |
2991 | if (!bio) | |
2992 | goto free_and_out; | |
2993 | ||
2994 | if (bio_ctr && bio_ctr(bio, bio_src, data)) | |
2995 | goto free_and_out; | |
2996 | ||
2997 | if (rq->bio) { | |
2998 | rq->biotail->bi_next = bio; | |
2999 | rq->biotail = bio; | |
3000 | } else { | |
3001 | rq->bio = rq->biotail = bio; | |
3002 | } | |
3003 | bio = NULL; | |
3004 | } | |
3005 | ||
3006 | /* Copy attributes of the original request to the clone request. */ | |
3007 | rq->__sector = blk_rq_pos(rq_src); | |
3008 | rq->__data_len = blk_rq_bytes(rq_src); | |
3009 | if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { | |
3010 | rq->rq_flags |= RQF_SPECIAL_PAYLOAD; | |
3011 | rq->special_vec = rq_src->special_vec; | |
3012 | } | |
3013 | rq->nr_phys_segments = rq_src->nr_phys_segments; | |
3014 | rq->ioprio = rq_src->ioprio; | |
3015 | ||
3016 | if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) | |
3017 | goto free_and_out; | |
3018 | ||
3019 | return 0; | |
3020 | ||
3021 | free_and_out: | |
3022 | if (bio) | |
3023 | bio_put(bio); | |
3024 | blk_rq_unprep_clone(rq); | |
3025 | ||
3026 | return -ENOMEM; | |
3027 | } | |
3028 | EXPORT_SYMBOL_GPL(blk_rq_prep_clone); | |
3029 | ||
f2b8f3ce CH |
3030 | /* |
3031 | * Steal bios from a request and add them to a bio list. | |
3032 | * The request must not have been partially completed before. | |
3033 | */ | |
3034 | void blk_steal_bios(struct bio_list *list, struct request *rq) | |
3035 | { | |
3036 | if (rq->bio) { | |
3037 | if (list->tail) | |
3038 | list->tail->bi_next = rq->bio; | |
3039 | else | |
3040 | list->head = rq->bio; | |
3041 | list->tail = rq->biotail; | |
3042 | ||
3043 | rq->bio = NULL; | |
3044 | rq->biotail = NULL; | |
3045 | } | |
3046 | ||
3047 | rq->__data_len = 0; | |
3048 | } | |
3049 | EXPORT_SYMBOL_GPL(blk_steal_bios); | |
3050 | ||
bd63141d ML |
3051 | static size_t order_to_size(unsigned int order) |
3052 | { | |
3053 | return (size_t)PAGE_SIZE << order; | |
3054 | } | |
3055 | ||
3056 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
3057 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
3058 | struct blk_mq_tags *tags) | |
bd63141d | 3059 | { |
bd63141d ML |
3060 | struct page *page; |
3061 | unsigned long flags; | |
3062 | ||
4f245d5b JG |
3063 | /* There is no need to clear a driver tags own mapping */ |
3064 | if (drv_tags == tags) | |
3065 | return; | |
3066 | ||
bd63141d ML |
3067 | list_for_each_entry(page, &tags->page_list, lru) { |
3068 | unsigned long start = (unsigned long)page_address(page); | |
3069 | unsigned long end = start + order_to_size(page->private); | |
3070 | int i; | |
3071 | ||
f32e4eaf | 3072 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
3073 | struct request *rq = drv_tags->rqs[i]; |
3074 | unsigned long rq_addr = (unsigned long)rq; | |
3075 | ||
3076 | if (rq_addr >= start && rq_addr < end) { | |
3077 | WARN_ON_ONCE(refcount_read(&rq->ref) != 0); | |
3078 | cmpxchg(&drv_tags->rqs[i], rq, NULL); | |
3079 | } | |
3080 | } | |
3081 | } | |
3082 | ||
3083 | /* | |
3084 | * Wait until all pending iteration is done. | |
3085 | * | |
3086 | * Request reference is cleared and it is guaranteed to be observed | |
3087 | * after the ->lock is released. | |
3088 | */ | |
3089 | spin_lock_irqsave(&drv_tags->lock, flags); | |
3090 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
3091 | } | |
3092 | ||
cc71a6f4 JA |
3093 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
3094 | unsigned int hctx_idx) | |
95363efd | 3095 | { |
f32e4eaf | 3096 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 3097 | struct page *page; |
320ae51f | 3098 | |
079a2e3e JG |
3099 | if (blk_mq_is_shared_tags(set->flags)) |
3100 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
3101 | else |
3102 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 3103 | |
65de57bb | 3104 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 3105 | int i; |
320ae51f | 3106 | |
24d2f903 | 3107 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
3108 | struct request *rq = tags->static_rqs[i]; |
3109 | ||
3110 | if (!rq) | |
e9b267d9 | 3111 | continue; |
d6296d39 | 3112 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 3113 | tags->static_rqs[i] = NULL; |
e9b267d9 | 3114 | } |
320ae51f | 3115 | } |
320ae51f | 3116 | |
f32e4eaf | 3117 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 3118 | |
24d2f903 CH |
3119 | while (!list_empty(&tags->page_list)) { |
3120 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 3121 | list_del_init(&page->lru); |
f75782e4 CM |
3122 | /* |
3123 | * Remove kmemleak object previously allocated in | |
273938bf | 3124 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
3125 | */ |
3126 | kmemleak_free(page_address(page)); | |
320ae51f JA |
3127 | __free_pages(page, page->private); |
3128 | } | |
cc71a6f4 | 3129 | } |
320ae51f | 3130 | |
e155b0c2 | 3131 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 3132 | { |
24d2f903 | 3133 | kfree(tags->rqs); |
cc71a6f4 | 3134 | tags->rqs = NULL; |
2af8cbe3 JA |
3135 | kfree(tags->static_rqs); |
3136 | tags->static_rqs = NULL; | |
320ae51f | 3137 | |
e155b0c2 | 3138 | blk_mq_free_tags(tags); |
320ae51f JA |
3139 | } |
3140 | ||
63064be1 JG |
3141 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
3142 | unsigned int hctx_idx, | |
3143 | unsigned int nr_tags, | |
e155b0c2 | 3144 | unsigned int reserved_tags) |
320ae51f | 3145 | { |
24d2f903 | 3146 | struct blk_mq_tags *tags; |
59f082e4 | 3147 | int node; |
320ae51f | 3148 | |
7d76f856 | 3149 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
3150 | if (node == NUMA_NO_NODE) |
3151 | node = set->numa_node; | |
3152 | ||
e155b0c2 JG |
3153 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
3154 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
3155 | if (!tags) |
3156 | return NULL; | |
320ae51f | 3157 | |
590b5b7d | 3158 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 3159 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 3160 | node); |
24d2f903 | 3161 | if (!tags->rqs) { |
e155b0c2 | 3162 | blk_mq_free_tags(tags); |
24d2f903 CH |
3163 | return NULL; |
3164 | } | |
320ae51f | 3165 | |
590b5b7d KC |
3166 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
3167 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
3168 | node); | |
2af8cbe3 JA |
3169 | if (!tags->static_rqs) { |
3170 | kfree(tags->rqs); | |
e155b0c2 | 3171 | blk_mq_free_tags(tags); |
2af8cbe3 JA |
3172 | return NULL; |
3173 | } | |
3174 | ||
cc71a6f4 JA |
3175 | return tags; |
3176 | } | |
3177 | ||
1d9bd516 TH |
3178 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
3179 | unsigned int hctx_idx, int node) | |
3180 | { | |
3181 | int ret; | |
3182 | ||
3183 | if (set->ops->init_request) { | |
3184 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
3185 | if (ret) | |
3186 | return ret; | |
3187 | } | |
3188 | ||
12f5b931 | 3189 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
3190 | return 0; |
3191 | } | |
3192 | ||
63064be1 JG |
3193 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
3194 | struct blk_mq_tags *tags, | |
3195 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
3196 | { |
3197 | unsigned int i, j, entries_per_page, max_order = 4; | |
3198 | size_t rq_size, left; | |
59f082e4 SL |
3199 | int node; |
3200 | ||
7d76f856 | 3201 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
3202 | if (node == NUMA_NO_NODE) |
3203 | node = set->numa_node; | |
cc71a6f4 JA |
3204 | |
3205 | INIT_LIST_HEAD(&tags->page_list); | |
3206 | ||
320ae51f JA |
3207 | /* |
3208 | * rq_size is the size of the request plus driver payload, rounded | |
3209 | * to the cacheline size | |
3210 | */ | |
24d2f903 | 3211 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 3212 | cache_line_size()); |
cc71a6f4 | 3213 | left = rq_size * depth; |
320ae51f | 3214 | |
cc71a6f4 | 3215 | for (i = 0; i < depth; ) { |
320ae51f JA |
3216 | int this_order = max_order; |
3217 | struct page *page; | |
3218 | int to_do; | |
3219 | void *p; | |
3220 | ||
b3a834b1 | 3221 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
3222 | this_order--; |
3223 | ||
3224 | do { | |
59f082e4 | 3225 | page = alloc_pages_node(node, |
36e1f3d1 | 3226 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 3227 | this_order); |
320ae51f JA |
3228 | if (page) |
3229 | break; | |
3230 | if (!this_order--) | |
3231 | break; | |
3232 | if (order_to_size(this_order) < rq_size) | |
3233 | break; | |
3234 | } while (1); | |
3235 | ||
3236 | if (!page) | |
24d2f903 | 3237 | goto fail; |
320ae51f JA |
3238 | |
3239 | page->private = this_order; | |
24d2f903 | 3240 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
3241 | |
3242 | p = page_address(page); | |
f75782e4 CM |
3243 | /* |
3244 | * Allow kmemleak to scan these pages as they contain pointers | |
3245 | * to additional allocations like via ops->init_request(). | |
3246 | */ | |
36e1f3d1 | 3247 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 3248 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 3249 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
3250 | left -= to_do * rq_size; |
3251 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
3252 | struct request *rq = p; |
3253 | ||
3254 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
3255 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
3256 | tags->static_rqs[i] = NULL; | |
3257 | goto fail; | |
e9b267d9 CH |
3258 | } |
3259 | ||
320ae51f JA |
3260 | p += rq_size; |
3261 | i++; | |
3262 | } | |
3263 | } | |
cc71a6f4 | 3264 | return 0; |
320ae51f | 3265 | |
24d2f903 | 3266 | fail: |
cc71a6f4 JA |
3267 | blk_mq_free_rqs(set, tags, hctx_idx); |
3268 | return -ENOMEM; | |
320ae51f JA |
3269 | } |
3270 | ||
bf0beec0 ML |
3271 | struct rq_iter_data { |
3272 | struct blk_mq_hw_ctx *hctx; | |
3273 | bool has_rq; | |
3274 | }; | |
3275 | ||
3276 | static bool blk_mq_has_request(struct request *rq, void *data, bool reserved) | |
3277 | { | |
3278 | struct rq_iter_data *iter_data = data; | |
3279 | ||
3280 | if (rq->mq_hctx != iter_data->hctx) | |
3281 | return true; | |
3282 | iter_data->has_rq = true; | |
3283 | return false; | |
3284 | } | |
3285 | ||
3286 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
3287 | { | |
3288 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
3289 | hctx->sched_tags : hctx->tags; | |
3290 | struct rq_iter_data data = { | |
3291 | .hctx = hctx, | |
3292 | }; | |
3293 | ||
3294 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
3295 | return data.has_rq; | |
3296 | } | |
3297 | ||
3298 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
3299 | struct blk_mq_hw_ctx *hctx) | |
3300 | { | |
3301 | if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu) | |
3302 | return false; | |
3303 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
3304 | return false; | |
3305 | return true; | |
3306 | } | |
3307 | ||
3308 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
3309 | { | |
3310 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3311 | struct blk_mq_hw_ctx, cpuhp_online); | |
3312 | ||
3313 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
3314 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
3315 | return 0; | |
3316 | ||
3317 | /* | |
3318 | * Prevent new request from being allocated on the current hctx. | |
3319 | * | |
3320 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
3321 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
3322 | * seen once we return from the tag allocator. | |
3323 | */ | |
3324 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3325 | smp_mb__after_atomic(); | |
3326 | ||
3327 | /* | |
3328 | * Try to grab a reference to the queue and wait for any outstanding | |
3329 | * requests. If we could not grab a reference the queue has been | |
3330 | * frozen and there are no requests. | |
3331 | */ | |
3332 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
3333 | while (blk_mq_hctx_has_requests(hctx)) | |
3334 | msleep(5); | |
3335 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
3336 | } | |
3337 | ||
3338 | return 0; | |
3339 | } | |
3340 | ||
3341 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
3342 | { | |
3343 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
3344 | struct blk_mq_hw_ctx, cpuhp_online); | |
3345 | ||
3346 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
3347 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
3348 | return 0; | |
3349 | } | |
3350 | ||
e57690fe JA |
3351 | /* |
3352 | * 'cpu' is going away. splice any existing rq_list entries from this | |
3353 | * software queue to the hw queue dispatch list, and ensure that it | |
3354 | * gets run. | |
3355 | */ | |
9467f859 | 3356 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 3357 | { |
9467f859 | 3358 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
3359 | struct blk_mq_ctx *ctx; |
3360 | LIST_HEAD(tmp); | |
c16d6b5a | 3361 | enum hctx_type type; |
484b4061 | 3362 | |
9467f859 | 3363 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
3364 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
3365 | return 0; | |
3366 | ||
e57690fe | 3367 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 3368 | type = hctx->type; |
484b4061 JA |
3369 | |
3370 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
3371 | if (!list_empty(&ctx->rq_lists[type])) { |
3372 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
3373 | blk_mq_hctx_clear_pending(hctx, ctx); |
3374 | } | |
3375 | spin_unlock(&ctx->lock); | |
3376 | ||
3377 | if (list_empty(&tmp)) | |
9467f859 | 3378 | return 0; |
484b4061 | 3379 | |
e57690fe JA |
3380 | spin_lock(&hctx->lock); |
3381 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
3382 | spin_unlock(&hctx->lock); | |
484b4061 JA |
3383 | |
3384 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 3385 | return 0; |
484b4061 JA |
3386 | } |
3387 | ||
9467f859 | 3388 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 3389 | { |
bf0beec0 ML |
3390 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3391 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3392 | &hctx->cpuhp_online); | |
9467f859 TG |
3393 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
3394 | &hctx->cpuhp_dead); | |
484b4061 JA |
3395 | } |
3396 | ||
364b6181 ML |
3397 | /* |
3398 | * Before freeing hw queue, clearing the flush request reference in | |
3399 | * tags->rqs[] for avoiding potential UAF. | |
3400 | */ | |
3401 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
3402 | unsigned int queue_depth, struct request *flush_rq) | |
3403 | { | |
3404 | int i; | |
3405 | unsigned long flags; | |
3406 | ||
3407 | /* The hw queue may not be mapped yet */ | |
3408 | if (!tags) | |
3409 | return; | |
3410 | ||
3411 | WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0); | |
3412 | ||
3413 | for (i = 0; i < queue_depth; i++) | |
3414 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
3415 | ||
3416 | /* | |
3417 | * Wait until all pending iteration is done. | |
3418 | * | |
3419 | * Request reference is cleared and it is guaranteed to be observed | |
3420 | * after the ->lock is released. | |
3421 | */ | |
3422 | spin_lock_irqsave(&tags->lock, flags); | |
3423 | spin_unlock_irqrestore(&tags->lock, flags); | |
3424 | } | |
3425 | ||
c3b4afca | 3426 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
3427 | static void blk_mq_exit_hctx(struct request_queue *q, |
3428 | struct blk_mq_tag_set *set, | |
3429 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
3430 | { | |
364b6181 ML |
3431 | struct request *flush_rq = hctx->fq->flush_rq; |
3432 | ||
8ab0b7dc ML |
3433 | if (blk_mq_hw_queue_mapped(hctx)) |
3434 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 3435 | |
364b6181 ML |
3436 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], |
3437 | set->queue_depth, flush_rq); | |
f70ced09 | 3438 | if (set->ops->exit_request) |
364b6181 | 3439 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 3440 | |
08e98fc6 ML |
3441 | if (set->ops->exit_hctx) |
3442 | set->ops->exit_hctx(hctx, hctx_idx); | |
3443 | ||
9467f859 | 3444 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
3445 | |
3446 | spin_lock(&q->unused_hctx_lock); | |
3447 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
3448 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
3449 | } |
3450 | ||
624dbe47 ML |
3451 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
3452 | struct blk_mq_tag_set *set, int nr_queue) | |
3453 | { | |
3454 | struct blk_mq_hw_ctx *hctx; | |
3455 | unsigned int i; | |
3456 | ||
3457 | queue_for_each_hw_ctx(q, hctx, i) { | |
3458 | if (i == nr_queue) | |
3459 | break; | |
477e19de | 3460 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 3461 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 3462 | } |
624dbe47 ML |
3463 | } |
3464 | ||
7c6c5b7c ML |
3465 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
3466 | { | |
3467 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
3468 | ||
3469 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
3470 | __alignof__(struct blk_mq_hw_ctx)) != | |
3471 | sizeof(struct blk_mq_hw_ctx)); | |
3472 | ||
3473 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
3474 | hw_ctx_size += sizeof(struct srcu_struct); | |
3475 | ||
3476 | return hw_ctx_size; | |
3477 | } | |
3478 | ||
08e98fc6 ML |
3479 | static int blk_mq_init_hctx(struct request_queue *q, |
3480 | struct blk_mq_tag_set *set, | |
3481 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 3482 | { |
7c6c5b7c ML |
3483 | hctx->queue_num = hctx_idx; |
3484 | ||
bf0beec0 ML |
3485 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
3486 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
3487 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
3488 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
3489 | ||
3490 | hctx->tags = set->tags[hctx_idx]; | |
3491 | ||
3492 | if (set->ops->init_hctx && | |
3493 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
3494 | goto unregister_cpu_notifier; | |
08e98fc6 | 3495 | |
7c6c5b7c ML |
3496 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
3497 | hctx->numa_node)) | |
3498 | goto exit_hctx; | |
3499 | return 0; | |
3500 | ||
3501 | exit_hctx: | |
3502 | if (set->ops->exit_hctx) | |
3503 | set->ops->exit_hctx(hctx, hctx_idx); | |
3504 | unregister_cpu_notifier: | |
3505 | blk_mq_remove_cpuhp(hctx); | |
3506 | return -1; | |
3507 | } | |
3508 | ||
3509 | static struct blk_mq_hw_ctx * | |
3510 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
3511 | int node) | |
3512 | { | |
3513 | struct blk_mq_hw_ctx *hctx; | |
3514 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
3515 | ||
3516 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
3517 | if (!hctx) | |
3518 | goto fail_alloc_hctx; | |
3519 | ||
3520 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
3521 | goto free_hctx; | |
3522 | ||
3523 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 3524 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
3525 | node = set->numa_node; |
3526 | hctx->numa_node = node; | |
08e98fc6 | 3527 | |
9f993737 | 3528 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
3529 | spin_lock_init(&hctx->lock); |
3530 | INIT_LIST_HEAD(&hctx->dispatch); | |
3531 | hctx->queue = q; | |
51db1c37 | 3532 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 3533 | |
2f8f1336 ML |
3534 | INIT_LIST_HEAD(&hctx->hctx_list); |
3535 | ||
320ae51f | 3536 | /* |
08e98fc6 ML |
3537 | * Allocate space for all possible cpus to avoid allocation at |
3538 | * runtime | |
320ae51f | 3539 | */ |
d904bfa7 | 3540 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 3541 | gfp, node); |
08e98fc6 | 3542 | if (!hctx->ctxs) |
7c6c5b7c | 3543 | goto free_cpumask; |
320ae51f | 3544 | |
5b202853 | 3545 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 3546 | gfp, node, false, false)) |
08e98fc6 | 3547 | goto free_ctxs; |
08e98fc6 | 3548 | hctx->nr_ctx = 0; |
320ae51f | 3549 | |
5815839b | 3550 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
3551 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
3552 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
3553 | ||
754a1572 | 3554 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 3555 | if (!hctx->fq) |
7c6c5b7c | 3556 | goto free_bitmap; |
320ae51f | 3557 | |
6a83e74d | 3558 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 3559 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 3560 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 3561 | |
7c6c5b7c | 3562 | return hctx; |
320ae51f | 3563 | |
08e98fc6 | 3564 | free_bitmap: |
88459642 | 3565 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
3566 | free_ctxs: |
3567 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
3568 | free_cpumask: |
3569 | free_cpumask_var(hctx->cpumask); | |
3570 | free_hctx: | |
3571 | kfree(hctx); | |
3572 | fail_alloc_hctx: | |
3573 | return NULL; | |
08e98fc6 | 3574 | } |
320ae51f | 3575 | |
320ae51f JA |
3576 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
3577 | unsigned int nr_hw_queues) | |
3578 | { | |
b3c661b1 JA |
3579 | struct blk_mq_tag_set *set = q->tag_set; |
3580 | unsigned int i, j; | |
320ae51f JA |
3581 | |
3582 | for_each_possible_cpu(i) { | |
3583 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
3584 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 3585 | int k; |
320ae51f | 3586 | |
320ae51f JA |
3587 | __ctx->cpu = i; |
3588 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
3589 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
3590 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
3591 | ||
320ae51f JA |
3592 | __ctx->queue = q; |
3593 | ||
320ae51f JA |
3594 | /* |
3595 | * Set local node, IFF we have more than one hw queue. If | |
3596 | * not, we remain on the home node of the device | |
3597 | */ | |
b3c661b1 JA |
3598 | for (j = 0; j < set->nr_maps; j++) { |
3599 | hctx = blk_mq_map_queue_type(q, j, i); | |
3600 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 3601 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 3602 | } |
320ae51f JA |
3603 | } |
3604 | } | |
3605 | ||
63064be1 JG |
3606 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3607 | unsigned int hctx_idx, | |
3608 | unsigned int depth) | |
cc71a6f4 | 3609 | { |
63064be1 JG |
3610 | struct blk_mq_tags *tags; |
3611 | int ret; | |
cc71a6f4 | 3612 | |
e155b0c2 | 3613 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
3614 | if (!tags) |
3615 | return NULL; | |
cc71a6f4 | 3616 | |
63064be1 JG |
3617 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
3618 | if (ret) { | |
e155b0c2 | 3619 | blk_mq_free_rq_map(tags); |
63064be1 JG |
3620 | return NULL; |
3621 | } | |
cc71a6f4 | 3622 | |
63064be1 | 3623 | return tags; |
cc71a6f4 JA |
3624 | } |
3625 | ||
63064be1 JG |
3626 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3627 | int hctx_idx) | |
cc71a6f4 | 3628 | { |
079a2e3e JG |
3629 | if (blk_mq_is_shared_tags(set->flags)) { |
3630 | set->tags[hctx_idx] = set->shared_tags; | |
1c0706a7 | 3631 | |
e155b0c2 | 3632 | return true; |
bd166ef1 | 3633 | } |
e155b0c2 | 3634 | |
63064be1 JG |
3635 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
3636 | set->queue_depth); | |
3637 | ||
3638 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
3639 | } |
3640 | ||
645db34e JG |
3641 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3642 | struct blk_mq_tags *tags, | |
3643 | unsigned int hctx_idx) | |
cc71a6f4 | 3644 | { |
645db34e JG |
3645 | if (tags) { |
3646 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 3647 | blk_mq_free_rq_map(tags); |
bd166ef1 | 3648 | } |
cc71a6f4 JA |
3649 | } |
3650 | ||
e155b0c2 JG |
3651 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3652 | unsigned int hctx_idx) | |
3653 | { | |
079a2e3e | 3654 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
3655 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
3656 | ||
3657 | set->tags[hctx_idx] = NULL; | |
cc71a6f4 JA |
3658 | } |
3659 | ||
4b855ad3 | 3660 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 3661 | { |
b3c661b1 | 3662 | unsigned int i, j, hctx_idx; |
320ae51f JA |
3663 | struct blk_mq_hw_ctx *hctx; |
3664 | struct blk_mq_ctx *ctx; | |
2a34c087 | 3665 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
3666 | |
3667 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 3668 | cpumask_clear(hctx->cpumask); |
320ae51f | 3669 | hctx->nr_ctx = 0; |
d416c92c | 3670 | hctx->dispatch_from = NULL; |
320ae51f JA |
3671 | } |
3672 | ||
3673 | /* | |
4b855ad3 | 3674 | * Map software to hardware queues. |
4412efec ML |
3675 | * |
3676 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 3677 | */ |
20e4d813 | 3678 | for_each_possible_cpu(i) { |
4412efec | 3679 | |
897bb0c7 | 3680 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 3681 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
3682 | if (!set->map[j].nr_queues) { |
3683 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3684 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 3685 | continue; |
bb94aea1 | 3686 | } |
fd689871 ML |
3687 | hctx_idx = set->map[j].mq_map[i]; |
3688 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
3689 | if (!set->tags[hctx_idx] && | |
63064be1 | 3690 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
3691 | /* |
3692 | * If tags initialization fail for some hctx, | |
3693 | * that hctx won't be brought online. In this | |
3694 | * case, remap the current ctx to hctx[0] which | |
3695 | * is guaranteed to always have tags allocated | |
3696 | */ | |
3697 | set->map[j].mq_map[i] = 0; | |
3698 | } | |
e5edd5f2 | 3699 | |
b3c661b1 | 3700 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3701 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3702 | /* |
3703 | * If the CPU is already set in the mask, then we've | |
3704 | * mapped this one already. This can happen if | |
3705 | * devices share queues across queue maps. | |
3706 | */ | |
3707 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3708 | continue; | |
3709 | ||
3710 | cpumask_set_cpu(i, hctx->cpumask); | |
3711 | hctx->type = j; | |
3712 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3713 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3714 | ||
3715 | /* | |
3716 | * If the nr_ctx type overflows, we have exceeded the | |
3717 | * amount of sw queues we can support. | |
3718 | */ | |
3719 | BUG_ON(!hctx->nr_ctx); | |
3720 | } | |
bb94aea1 JW |
3721 | |
3722 | for (; j < HCTX_MAX_TYPES; j++) | |
3723 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3724 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3725 | } |
506e931f JA |
3726 | |
3727 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
3728 | /* |
3729 | * If no software queues are mapped to this hardware queue, | |
3730 | * disable it and free the request entries. | |
3731 | */ | |
3732 | if (!hctx->nr_ctx) { | |
3733 | /* Never unmap queue 0. We need it as a | |
3734 | * fallback in case of a new remap fails | |
3735 | * allocation | |
3736 | */ | |
e155b0c2 JG |
3737 | if (i) |
3738 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3739 | |
3740 | hctx->tags = NULL; | |
3741 | continue; | |
3742 | } | |
484b4061 | 3743 | |
2a34c087 ML |
3744 | hctx->tags = set->tags[i]; |
3745 | WARN_ON(!hctx->tags); | |
3746 | ||
889fa31f CY |
3747 | /* |
3748 | * Set the map size to the number of mapped software queues. | |
3749 | * This is more accurate and more efficient than looping | |
3750 | * over all possibly mapped software queues. | |
3751 | */ | |
88459642 | 3752 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3753 | |
484b4061 JA |
3754 | /* |
3755 | * Initialize batch roundrobin counts | |
3756 | */ | |
f82ddf19 | 3757 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
3758 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
3759 | } | |
320ae51f JA |
3760 | } |
3761 | ||
8e8320c9 JA |
3762 | /* |
3763 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
3764 | * the queue isn't live yet. | |
3765 | */ | |
2404e607 | 3766 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
3767 | { |
3768 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
3769 | int i; |
3770 | ||
2404e607 | 3771 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3772 | if (shared) { |
51db1c37 | 3773 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3774 | } else { |
3775 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3776 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3777 | } |
2404e607 JM |
3778 | } |
3779 | } | |
3780 | ||
655ac300 HR |
3781 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3782 | bool shared) | |
2404e607 JM |
3783 | { |
3784 | struct request_queue *q; | |
0d2602ca | 3785 | |
705cda97 BVA |
3786 | lockdep_assert_held(&set->tag_list_lock); |
3787 | ||
0d2602ca JA |
3788 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3789 | blk_mq_freeze_queue(q); | |
2404e607 | 3790 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3791 | blk_mq_unfreeze_queue(q); |
3792 | } | |
3793 | } | |
3794 | ||
3795 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3796 | { | |
3797 | struct blk_mq_tag_set *set = q->tag_set; | |
3798 | ||
0d2602ca | 3799 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3800 | list_del(&q->tag_set_list); |
2404e607 JM |
3801 | if (list_is_singular(&set->tag_list)) { |
3802 | /* just transitioned to unshared */ | |
51db1c37 | 3803 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3804 | /* update existing queue */ |
655ac300 | 3805 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3806 | } |
0d2602ca | 3807 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3808 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3809 | } |
3810 | ||
3811 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3812 | struct request_queue *q) | |
3813 | { | |
0d2602ca | 3814 | mutex_lock(&set->tag_list_lock); |
2404e607 | 3815 | |
ff821d27 JA |
3816 | /* |
3817 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
3818 | */ | |
3819 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
3820 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
3821 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 3822 | /* update existing queue */ |
655ac300 | 3823 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 3824 | } |
51db1c37 | 3825 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 3826 | queue_set_hctx_shared(q, true); |
08c875cb | 3827 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 3828 | |
0d2602ca JA |
3829 | mutex_unlock(&set->tag_list_lock); |
3830 | } | |
3831 | ||
1db4909e ML |
3832 | /* All allocations will be freed in release handler of q->mq_kobj */ |
3833 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
3834 | { | |
3835 | struct blk_mq_ctxs *ctxs; | |
3836 | int cpu; | |
3837 | ||
3838 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
3839 | if (!ctxs) | |
3840 | return -ENOMEM; | |
3841 | ||
3842 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
3843 | if (!ctxs->queue_ctx) | |
3844 | goto fail; | |
3845 | ||
3846 | for_each_possible_cpu(cpu) { | |
3847 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
3848 | ctx->ctxs = ctxs; | |
3849 | } | |
3850 | ||
3851 | q->mq_kobj = &ctxs->kobj; | |
3852 | q->queue_ctx = ctxs->queue_ctx; | |
3853 | ||
3854 | return 0; | |
3855 | fail: | |
3856 | kfree(ctxs); | |
3857 | return -ENOMEM; | |
3858 | } | |
3859 | ||
e09aae7e ML |
3860 | /* |
3861 | * It is the actual release handler for mq, but we do it from | |
3862 | * request queue's release handler for avoiding use-after-free | |
3863 | * and headache because q->mq_kobj shouldn't have been introduced, | |
3864 | * but we can't group ctx/kctx kobj without it. | |
3865 | */ | |
3866 | void blk_mq_release(struct request_queue *q) | |
3867 | { | |
2f8f1336 ML |
3868 | struct blk_mq_hw_ctx *hctx, *next; |
3869 | int i; | |
e09aae7e | 3870 | |
2f8f1336 ML |
3871 | queue_for_each_hw_ctx(q, hctx, i) |
3872 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
3873 | ||
3874 | /* all hctx are in .unused_hctx_list now */ | |
3875 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
3876 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 3877 | kobject_put(&hctx->kobj); |
c3b4afca | 3878 | } |
e09aae7e ML |
3879 | |
3880 | kfree(q->queue_hw_ctx); | |
3881 | ||
7ea5fe31 ML |
3882 | /* |
3883 | * release .mq_kobj and sw queue's kobject now because | |
3884 | * both share lifetime with request queue. | |
3885 | */ | |
3886 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
3887 | } |
3888 | ||
5ec780a6 | 3889 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 3890 | void *queuedata) |
b62c21b7 | 3891 | { |
26a9750a CH |
3892 | struct request_queue *q; |
3893 | int ret; | |
b62c21b7 | 3894 | |
26a9750a CH |
3895 | q = blk_alloc_queue(set->numa_node); |
3896 | if (!q) | |
b62c21b7 | 3897 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
3898 | q->queuedata = queuedata; |
3899 | ret = blk_mq_init_allocated_queue(set, q); | |
3900 | if (ret) { | |
3901 | blk_cleanup_queue(q); | |
3902 | return ERR_PTR(ret); | |
3903 | } | |
b62c21b7 MS |
3904 | return q; |
3905 | } | |
2f227bb9 CH |
3906 | |
3907 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
3908 | { | |
3909 | return blk_mq_init_queue_data(set, NULL); | |
3910 | } | |
b62c21b7 MS |
3911 | EXPORT_SYMBOL(blk_mq_init_queue); |
3912 | ||
4dcc4874 CH |
3913 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
3914 | struct lock_class_key *lkclass) | |
9316a9ed JA |
3915 | { |
3916 | struct request_queue *q; | |
b461dfc4 | 3917 | struct gendisk *disk; |
9316a9ed | 3918 | |
b461dfc4 CH |
3919 | q = blk_mq_init_queue_data(set, queuedata); |
3920 | if (IS_ERR(q)) | |
3921 | return ERR_CAST(q); | |
9316a9ed | 3922 | |
4a1fa41d | 3923 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 CH |
3924 | if (!disk) { |
3925 | blk_cleanup_queue(q); | |
3926 | return ERR_PTR(-ENOMEM); | |
9316a9ed | 3927 | } |
b461dfc4 | 3928 | return disk; |
9316a9ed | 3929 | } |
b461dfc4 | 3930 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 3931 | |
34d11ffa JW |
3932 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
3933 | struct blk_mq_tag_set *set, struct request_queue *q, | |
3934 | int hctx_idx, int node) | |
3935 | { | |
2f8f1336 | 3936 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 3937 | |
2f8f1336 ML |
3938 | /* reuse dead hctx first */ |
3939 | spin_lock(&q->unused_hctx_lock); | |
3940 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
3941 | if (tmp->numa_node == node) { | |
3942 | hctx = tmp; | |
3943 | break; | |
3944 | } | |
3945 | } | |
3946 | if (hctx) | |
3947 | list_del_init(&hctx->hctx_list); | |
3948 | spin_unlock(&q->unused_hctx_lock); | |
3949 | ||
3950 | if (!hctx) | |
3951 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 3952 | if (!hctx) |
7c6c5b7c | 3953 | goto fail; |
34d11ffa | 3954 | |
7c6c5b7c ML |
3955 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
3956 | goto free_hctx; | |
34d11ffa JW |
3957 | |
3958 | return hctx; | |
7c6c5b7c ML |
3959 | |
3960 | free_hctx: | |
3961 | kobject_put(&hctx->kobj); | |
3962 | fail: | |
3963 | return NULL; | |
34d11ffa JW |
3964 | } |
3965 | ||
868f2f0b KB |
3966 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
3967 | struct request_queue *q) | |
320ae51f | 3968 | { |
e01ad46d | 3969 | int i, j, end; |
868f2f0b | 3970 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 3971 | |
ac0d6b92 BVA |
3972 | if (q->nr_hw_queues < set->nr_hw_queues) { |
3973 | struct blk_mq_hw_ctx **new_hctxs; | |
3974 | ||
3975 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
3976 | sizeof(*new_hctxs), GFP_KERNEL, | |
3977 | set->numa_node); | |
3978 | if (!new_hctxs) | |
3979 | return; | |
3980 | if (hctxs) | |
3981 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
3982 | sizeof(*hctxs)); | |
3983 | q->queue_hw_ctx = new_hctxs; | |
ac0d6b92 BVA |
3984 | kfree(hctxs); |
3985 | hctxs = new_hctxs; | |
3986 | } | |
3987 | ||
fb350e0a ML |
3988 | /* protect against switching io scheduler */ |
3989 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 3990 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 3991 | int node; |
34d11ffa | 3992 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 3993 | |
7d76f856 | 3994 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
3995 | /* |
3996 | * If the hw queue has been mapped to another numa node, | |
3997 | * we need to realloc the hctx. If allocation fails, fallback | |
3998 | * to use the previous one. | |
3999 | */ | |
4000 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
4001 | continue; | |
868f2f0b | 4002 | |
34d11ffa JW |
4003 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
4004 | if (hctx) { | |
2f8f1336 | 4005 | if (hctxs[i]) |
34d11ffa | 4006 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
4007 | hctxs[i] = hctx; |
4008 | } else { | |
4009 | if (hctxs[i]) | |
4010 | pr_warn("Allocate new hctx on node %d fails,\ | |
4011 | fallback to previous one on node %d\n", | |
4012 | node, hctxs[i]->numa_node); | |
4013 | else | |
4014 | break; | |
868f2f0b | 4015 | } |
320ae51f | 4016 | } |
e01ad46d JW |
4017 | /* |
4018 | * Increasing nr_hw_queues fails. Free the newly allocated | |
4019 | * hctxs and keep the previous q->nr_hw_queues. | |
4020 | */ | |
4021 | if (i != set->nr_hw_queues) { | |
4022 | j = q->nr_hw_queues; | |
4023 | end = i; | |
4024 | } else { | |
4025 | j = i; | |
4026 | end = q->nr_hw_queues; | |
4027 | q->nr_hw_queues = set->nr_hw_queues; | |
4028 | } | |
34d11ffa | 4029 | |
e01ad46d | 4030 | for (; j < end; j++) { |
868f2f0b KB |
4031 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
4032 | ||
4033 | if (hctx) { | |
868f2f0b | 4034 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 4035 | hctxs[j] = NULL; |
868f2f0b KB |
4036 | } |
4037 | } | |
fb350e0a | 4038 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
4039 | } |
4040 | ||
26a9750a CH |
4041 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
4042 | struct request_queue *q) | |
868f2f0b | 4043 | { |
66841672 ML |
4044 | /* mark the queue as mq asap */ |
4045 | q->mq_ops = set->ops; | |
4046 | ||
34dbad5d | 4047 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
4048 | blk_mq_poll_stats_bkt, |
4049 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
4050 | if (!q->poll_cb) |
4051 | goto err_exit; | |
4052 | ||
1db4909e | 4053 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 4054 | goto err_poll; |
868f2f0b | 4055 | |
737f98cf ML |
4056 | /* init q->mq_kobj and sw queues' kobjects */ |
4057 | blk_mq_sysfs_init(q); | |
4058 | ||
2f8f1336 ML |
4059 | INIT_LIST_HEAD(&q->unused_hctx_list); |
4060 | spin_lock_init(&q->unused_hctx_lock); | |
4061 | ||
868f2f0b KB |
4062 | blk_mq_realloc_hw_ctxs(set, q); |
4063 | if (!q->nr_hw_queues) | |
4064 | goto err_hctxs; | |
320ae51f | 4065 | |
287922eb | 4066 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 4067 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 4068 | |
a8908939 | 4069 | q->tag_set = set; |
320ae51f | 4070 | |
94eddfbe | 4071 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
4072 | if (set->nr_maps > HCTX_TYPE_POLL && |
4073 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 4074 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 4075 | |
2849450a | 4076 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
4077 | INIT_LIST_HEAD(&q->requeue_list); |
4078 | spin_lock_init(&q->requeue_lock); | |
4079 | ||
eba71768 JA |
4080 | q->nr_requests = set->queue_depth; |
4081 | ||
64f1c21e JA |
4082 | /* |
4083 | * Default to classic polling | |
4084 | */ | |
29ece8b4 | 4085 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 4086 | |
24d2f903 | 4087 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 4088 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 4089 | blk_mq_map_swqueue(q); |
26a9750a | 4090 | return 0; |
18741986 | 4091 | |
320ae51f | 4092 | err_hctxs: |
868f2f0b | 4093 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 4094 | q->nr_hw_queues = 0; |
1db4909e | 4095 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
4096 | err_poll: |
4097 | blk_stat_free_callback(q->poll_cb); | |
4098 | q->poll_cb = NULL; | |
c7de5726 ML |
4099 | err_exit: |
4100 | q->mq_ops = NULL; | |
26a9750a | 4101 | return -ENOMEM; |
320ae51f | 4102 | } |
b62c21b7 | 4103 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 4104 | |
c7e2d94b ML |
4105 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
4106 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 4107 | { |
630ef623 | 4108 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 4109 | |
630ef623 | 4110 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 4111 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
4112 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
4113 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 4114 | } |
320ae51f | 4115 | |
a5164405 JA |
4116 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
4117 | { | |
4118 | int i; | |
4119 | ||
079a2e3e JG |
4120 | if (blk_mq_is_shared_tags(set->flags)) { |
4121 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
4122 | BLK_MQ_NO_HCTX_IDX, |
4123 | set->queue_depth); | |
079a2e3e | 4124 | if (!set->shared_tags) |
e155b0c2 JG |
4125 | return -ENOMEM; |
4126 | } | |
4127 | ||
8229cca8 | 4128 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 4129 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 4130 | goto out_unwind; |
8229cca8 XT |
4131 | cond_resched(); |
4132 | } | |
a5164405 JA |
4133 | |
4134 | return 0; | |
4135 | ||
4136 | out_unwind: | |
4137 | while (--i >= 0) | |
e155b0c2 JG |
4138 | __blk_mq_free_map_and_rqs(set, i); |
4139 | ||
079a2e3e JG |
4140 | if (blk_mq_is_shared_tags(set->flags)) { |
4141 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 4142 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 4143 | } |
a5164405 | 4144 | |
a5164405 JA |
4145 | return -ENOMEM; |
4146 | } | |
4147 | ||
4148 | /* | |
4149 | * Allocate the request maps associated with this tag_set. Note that this | |
4150 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
4151 | * will be updated to reflect the allocated depth. | |
4152 | */ | |
63064be1 | 4153 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
4154 | { |
4155 | unsigned int depth; | |
4156 | int err; | |
4157 | ||
4158 | depth = set->queue_depth; | |
4159 | do { | |
4160 | err = __blk_mq_alloc_rq_maps(set); | |
4161 | if (!err) | |
4162 | break; | |
4163 | ||
4164 | set->queue_depth >>= 1; | |
4165 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
4166 | err = -ENOMEM; | |
4167 | break; | |
4168 | } | |
4169 | } while (set->queue_depth); | |
4170 | ||
4171 | if (!set->queue_depth || err) { | |
4172 | pr_err("blk-mq: failed to allocate request map\n"); | |
4173 | return -ENOMEM; | |
4174 | } | |
4175 | ||
4176 | if (depth != set->queue_depth) | |
4177 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
4178 | depth, set->queue_depth); | |
4179 | ||
4180 | return 0; | |
4181 | } | |
4182 | ||
ebe8bddb OS |
4183 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
4184 | { | |
6e66b493 BVA |
4185 | /* |
4186 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
4187 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
4188 | * number of hardware queues. | |
4189 | */ | |
4190 | if (set->nr_maps == 1) | |
4191 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
4192 | ||
59388702 | 4193 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
4194 | int i; |
4195 | ||
7d4901a9 ML |
4196 | /* |
4197 | * transport .map_queues is usually done in the following | |
4198 | * way: | |
4199 | * | |
4200 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
4201 | * mask = get_cpu_mask(queue) | |
4202 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 4203 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
4204 | * } |
4205 | * | |
4206 | * When we need to remap, the table has to be cleared for | |
4207 | * killing stale mapping since one CPU may not be mapped | |
4208 | * to any hw queue. | |
4209 | */ | |
b3c661b1 JA |
4210 | for (i = 0; i < set->nr_maps; i++) |
4211 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 4212 | |
ebe8bddb | 4213 | return set->ops->map_queues(set); |
b3c661b1 JA |
4214 | } else { |
4215 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 4216 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 4217 | } |
ebe8bddb OS |
4218 | } |
4219 | ||
f7e76dbc BVA |
4220 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
4221 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
4222 | { | |
4223 | struct blk_mq_tags **new_tags; | |
4224 | ||
4225 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
4226 | return 0; | |
4227 | ||
4228 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
4229 | GFP_KERNEL, set->numa_node); | |
4230 | if (!new_tags) | |
4231 | return -ENOMEM; | |
4232 | ||
4233 | if (set->tags) | |
4234 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
4235 | sizeof(*set->tags)); | |
4236 | kfree(set->tags); | |
4237 | set->tags = new_tags; | |
4238 | set->nr_hw_queues = new_nr_hw_queues; | |
4239 | ||
4240 | return 0; | |
4241 | } | |
4242 | ||
91cdf265 MI |
4243 | static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set, |
4244 | int new_nr_hw_queues) | |
4245 | { | |
4246 | return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues); | |
4247 | } | |
4248 | ||
a4391c64 JA |
4249 | /* |
4250 | * Alloc a tag set to be associated with one or more request queues. | |
4251 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 4252 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
4253 | * value will be stored in set->queue_depth. |
4254 | */ | |
24d2f903 CH |
4255 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
4256 | { | |
b3c661b1 | 4257 | int i, ret; |
da695ba2 | 4258 | |
205fb5f5 BVA |
4259 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
4260 | ||
24d2f903 CH |
4261 | if (!set->nr_hw_queues) |
4262 | return -EINVAL; | |
a4391c64 | 4263 | if (!set->queue_depth) |
24d2f903 CH |
4264 | return -EINVAL; |
4265 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
4266 | return -EINVAL; | |
4267 | ||
7d7e0f90 | 4268 | if (!set->ops->queue_rq) |
24d2f903 CH |
4269 | return -EINVAL; |
4270 | ||
de148297 ML |
4271 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
4272 | return -EINVAL; | |
4273 | ||
a4391c64 JA |
4274 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
4275 | pr_info("blk-mq: reduced tag depth to %u\n", | |
4276 | BLK_MQ_MAX_DEPTH); | |
4277 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
4278 | } | |
24d2f903 | 4279 | |
b3c661b1 JA |
4280 | if (!set->nr_maps) |
4281 | set->nr_maps = 1; | |
4282 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
4283 | return -EINVAL; | |
4284 | ||
6637fadf SL |
4285 | /* |
4286 | * If a crashdump is active, then we are potentially in a very | |
4287 | * memory constrained environment. Limit us to 1 queue and | |
4288 | * 64 tags to prevent using too much memory. | |
4289 | */ | |
4290 | if (is_kdump_kernel()) { | |
4291 | set->nr_hw_queues = 1; | |
59388702 | 4292 | set->nr_maps = 1; |
6637fadf SL |
4293 | set->queue_depth = min(64U, set->queue_depth); |
4294 | } | |
868f2f0b | 4295 | /* |
392546ae JA |
4296 | * There is no use for more h/w queues than cpus if we just have |
4297 | * a single map | |
868f2f0b | 4298 | */ |
392546ae | 4299 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4300 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 4301 | |
91cdf265 | 4302 | if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0) |
a5164405 | 4303 | return -ENOMEM; |
24d2f903 | 4304 | |
da695ba2 | 4305 | ret = -ENOMEM; |
b3c661b1 JA |
4306 | for (i = 0; i < set->nr_maps; i++) { |
4307 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 4308 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
4309 | GFP_KERNEL, set->numa_node); |
4310 | if (!set->map[i].mq_map) | |
4311 | goto out_free_mq_map; | |
59388702 | 4312 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 4313 | } |
bdd17e75 | 4314 | |
ebe8bddb | 4315 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
4316 | if (ret) |
4317 | goto out_free_mq_map; | |
4318 | ||
63064be1 | 4319 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 4320 | if (ret) |
bdd17e75 | 4321 | goto out_free_mq_map; |
24d2f903 | 4322 | |
0d2602ca JA |
4323 | mutex_init(&set->tag_list_lock); |
4324 | INIT_LIST_HEAD(&set->tag_list); | |
4325 | ||
24d2f903 | 4326 | return 0; |
bdd17e75 CH |
4327 | |
4328 | out_free_mq_map: | |
b3c661b1 JA |
4329 | for (i = 0; i < set->nr_maps; i++) { |
4330 | kfree(set->map[i].mq_map); | |
4331 | set->map[i].mq_map = NULL; | |
4332 | } | |
5676e7b6 RE |
4333 | kfree(set->tags); |
4334 | set->tags = NULL; | |
da695ba2 | 4335 | return ret; |
24d2f903 CH |
4336 | } |
4337 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
4338 | ||
cdb14e0f CH |
4339 | /* allocate and initialize a tagset for a simple single-queue device */ |
4340 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
4341 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
4342 | unsigned int set_flags) | |
4343 | { | |
4344 | memset(set, 0, sizeof(*set)); | |
4345 | set->ops = ops; | |
4346 | set->nr_hw_queues = 1; | |
4347 | set->nr_maps = 1; | |
4348 | set->queue_depth = queue_depth; | |
4349 | set->numa_node = NUMA_NO_NODE; | |
4350 | set->flags = set_flags; | |
4351 | return blk_mq_alloc_tag_set(set); | |
4352 | } | |
4353 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
4354 | ||
24d2f903 CH |
4355 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
4356 | { | |
b3c661b1 | 4357 | int i, j; |
24d2f903 | 4358 | |
f7e76dbc | 4359 | for (i = 0; i < set->nr_hw_queues; i++) |
e155b0c2 | 4360 | __blk_mq_free_map_and_rqs(set, i); |
484b4061 | 4361 | |
079a2e3e JG |
4362 | if (blk_mq_is_shared_tags(set->flags)) { |
4363 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
4364 | BLK_MQ_NO_HCTX_IDX); |
4365 | } | |
32bc15af | 4366 | |
b3c661b1 JA |
4367 | for (j = 0; j < set->nr_maps; j++) { |
4368 | kfree(set->map[j].mq_map); | |
4369 | set->map[j].mq_map = NULL; | |
4370 | } | |
bdd17e75 | 4371 | |
981bd189 | 4372 | kfree(set->tags); |
5676e7b6 | 4373 | set->tags = NULL; |
24d2f903 CH |
4374 | } |
4375 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
4376 | ||
e3a2b3f9 JA |
4377 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
4378 | { | |
4379 | struct blk_mq_tag_set *set = q->tag_set; | |
4380 | struct blk_mq_hw_ctx *hctx; | |
4381 | int i, ret; | |
4382 | ||
bd166ef1 | 4383 | if (!set) |
e3a2b3f9 JA |
4384 | return -EINVAL; |
4385 | ||
e5fa8140 AZ |
4386 | if (q->nr_requests == nr) |
4387 | return 0; | |
4388 | ||
70f36b60 | 4389 | blk_mq_freeze_queue(q); |
24f5a90f | 4390 | blk_mq_quiesce_queue(q); |
70f36b60 | 4391 | |
e3a2b3f9 JA |
4392 | ret = 0; |
4393 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
4394 | if (!hctx->tags) |
4395 | continue; | |
bd166ef1 JA |
4396 | /* |
4397 | * If we're using an MQ scheduler, just update the scheduler | |
4398 | * queue depth. This is similar to what the old code would do. | |
4399 | */ | |
f6adcef5 | 4400 | if (hctx->sched_tags) { |
70f36b60 | 4401 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 4402 | nr, true); |
f6adcef5 JG |
4403 | } else { |
4404 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
4405 | false); | |
70f36b60 | 4406 | } |
e3a2b3f9 JA |
4407 | if (ret) |
4408 | break; | |
77f1e0a5 JA |
4409 | if (q->elevator && q->elevator->type->ops.depth_updated) |
4410 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 4411 | } |
d97e594c | 4412 | if (!ret) { |
e3a2b3f9 | 4413 | q->nr_requests = nr; |
079a2e3e | 4414 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 4415 | if (q->elevator) |
079a2e3e | 4416 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 4417 | else |
079a2e3e | 4418 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 4419 | } |
d97e594c | 4420 | } |
e3a2b3f9 | 4421 | |
24f5a90f | 4422 | blk_mq_unquiesce_queue(q); |
70f36b60 | 4423 | blk_mq_unfreeze_queue(q); |
70f36b60 | 4424 | |
e3a2b3f9 JA |
4425 | return ret; |
4426 | } | |
4427 | ||
d48ece20 JW |
4428 | /* |
4429 | * request_queue and elevator_type pair. | |
4430 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
4431 | * the elevator_type associated with a request_queue. | |
4432 | */ | |
4433 | struct blk_mq_qe_pair { | |
4434 | struct list_head node; | |
4435 | struct request_queue *q; | |
4436 | struct elevator_type *type; | |
4437 | }; | |
4438 | ||
4439 | /* | |
4440 | * Cache the elevator_type in qe pair list and switch the | |
4441 | * io scheduler to 'none' | |
4442 | */ | |
4443 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
4444 | struct request_queue *q) | |
4445 | { | |
4446 | struct blk_mq_qe_pair *qe; | |
4447 | ||
4448 | if (!q->elevator) | |
4449 | return true; | |
4450 | ||
4451 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
4452 | if (!qe) | |
4453 | return false; | |
4454 | ||
4455 | INIT_LIST_HEAD(&qe->node); | |
4456 | qe->q = q; | |
4457 | qe->type = q->elevator->type; | |
4458 | list_add(&qe->node, head); | |
4459 | ||
4460 | mutex_lock(&q->sysfs_lock); | |
4461 | /* | |
4462 | * After elevator_switch_mq, the previous elevator_queue will be | |
4463 | * released by elevator_release. The reference of the io scheduler | |
4464 | * module get by elevator_get will also be put. So we need to get | |
4465 | * a reference of the io scheduler module here to prevent it to be | |
4466 | * removed. | |
4467 | */ | |
4468 | __module_get(qe->type->elevator_owner); | |
4469 | elevator_switch_mq(q, NULL); | |
4470 | mutex_unlock(&q->sysfs_lock); | |
4471 | ||
4472 | return true; | |
4473 | } | |
4474 | ||
4475 | static void blk_mq_elv_switch_back(struct list_head *head, | |
4476 | struct request_queue *q) | |
4477 | { | |
4478 | struct blk_mq_qe_pair *qe; | |
4479 | struct elevator_type *t = NULL; | |
4480 | ||
4481 | list_for_each_entry(qe, head, node) | |
4482 | if (qe->q == q) { | |
4483 | t = qe->type; | |
4484 | break; | |
4485 | } | |
4486 | ||
4487 | if (!t) | |
4488 | return; | |
4489 | ||
4490 | list_del(&qe->node); | |
4491 | kfree(qe); | |
4492 | ||
4493 | mutex_lock(&q->sysfs_lock); | |
4494 | elevator_switch_mq(q, t); | |
4495 | mutex_unlock(&q->sysfs_lock); | |
4496 | } | |
4497 | ||
e4dc2b32 KB |
4498 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
4499 | int nr_hw_queues) | |
868f2f0b KB |
4500 | { |
4501 | struct request_queue *q; | |
d48ece20 | 4502 | LIST_HEAD(head); |
e01ad46d | 4503 | int prev_nr_hw_queues; |
868f2f0b | 4504 | |
705cda97 BVA |
4505 | lockdep_assert_held(&set->tag_list_lock); |
4506 | ||
392546ae | 4507 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 4508 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
4509 | if (nr_hw_queues < 1) |
4510 | return; | |
4511 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
4512 | return; |
4513 | ||
4514 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4515 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
4516 | /* |
4517 | * Switch IO scheduler to 'none', cleaning up the data associated | |
4518 | * with the previous scheduler. We will switch back once we are done | |
4519 | * updating the new sw to hw queue mappings. | |
4520 | */ | |
4521 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4522 | if (!blk_mq_elv_switch_none(&head, q)) | |
4523 | goto switch_back; | |
868f2f0b | 4524 | |
477e19de JW |
4525 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4526 | blk_mq_debugfs_unregister_hctxs(q); | |
4527 | blk_mq_sysfs_unregister(q); | |
4528 | } | |
4529 | ||
a2584e43 | 4530 | prev_nr_hw_queues = set->nr_hw_queues; |
f7e76dbc BVA |
4531 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
4532 | 0) | |
4533 | goto reregister; | |
4534 | ||
868f2f0b | 4535 | set->nr_hw_queues = nr_hw_queues; |
e01ad46d | 4536 | fallback: |
aa880ad6 | 4537 | blk_mq_update_queue_map(set); |
868f2f0b KB |
4538 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4539 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d | 4540 | if (q->nr_hw_queues != set->nr_hw_queues) { |
a846a8e6 YB |
4541 | int i = prev_nr_hw_queues; |
4542 | ||
e01ad46d JW |
4543 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", |
4544 | nr_hw_queues, prev_nr_hw_queues); | |
a846a8e6 YB |
4545 | for (; i < set->nr_hw_queues; i++) |
4546 | __blk_mq_free_map_and_rqs(set, i); | |
4547 | ||
e01ad46d | 4548 | set->nr_hw_queues = prev_nr_hw_queues; |
7d76f856 | 4549 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
4550 | goto fallback; |
4551 | } | |
477e19de JW |
4552 | blk_mq_map_swqueue(q); |
4553 | } | |
4554 | ||
f7e76dbc | 4555 | reregister: |
477e19de JW |
4556 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4557 | blk_mq_sysfs_register(q); | |
4558 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
4559 | } |
4560 | ||
d48ece20 JW |
4561 | switch_back: |
4562 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4563 | blk_mq_elv_switch_back(&head, q); | |
4564 | ||
868f2f0b KB |
4565 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
4566 | blk_mq_unfreeze_queue(q); | |
4567 | } | |
e4dc2b32 KB |
4568 | |
4569 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
4570 | { | |
4571 | mutex_lock(&set->tag_list_lock); | |
4572 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
4573 | mutex_unlock(&set->tag_list_lock); | |
4574 | } | |
868f2f0b KB |
4575 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
4576 | ||
34dbad5d OS |
4577 | /* Enable polling stats and return whether they were already enabled. */ |
4578 | static bool blk_poll_stats_enable(struct request_queue *q) | |
4579 | { | |
48b5c1fb | 4580 | if (q->poll_stat) |
34dbad5d | 4581 | return true; |
48b5c1fb JA |
4582 | |
4583 | return blk_stats_alloc_enable(q); | |
34dbad5d OS |
4584 | } |
4585 | ||
4586 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
4587 | { | |
4588 | /* | |
4589 | * We don't arm the callback if polling stats are not enabled or the | |
4590 | * callback is already active. | |
4591 | */ | |
48b5c1fb | 4592 | if (!q->poll_stat || blk_stat_is_active(q->poll_cb)) |
34dbad5d OS |
4593 | return; |
4594 | ||
4595 | blk_stat_activate_msecs(q->poll_cb, 100); | |
4596 | } | |
4597 | ||
4598 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
4599 | { | |
4600 | struct request_queue *q = cb->data; | |
720b8ccc | 4601 | int bucket; |
34dbad5d | 4602 | |
720b8ccc SB |
4603 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
4604 | if (cb->stat[bucket].nr_samples) | |
4605 | q->poll_stat[bucket] = cb->stat[bucket]; | |
4606 | } | |
34dbad5d OS |
4607 | } |
4608 | ||
64f1c21e | 4609 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
4610 | struct request *rq) |
4611 | { | |
64f1c21e | 4612 | unsigned long ret = 0; |
720b8ccc | 4613 | int bucket; |
64f1c21e JA |
4614 | |
4615 | /* | |
4616 | * If stats collection isn't on, don't sleep but turn it on for | |
4617 | * future users | |
4618 | */ | |
34dbad5d | 4619 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
4620 | return 0; |
4621 | ||
64f1c21e JA |
4622 | /* |
4623 | * As an optimistic guess, use half of the mean service time | |
4624 | * for this type of request. We can (and should) make this smarter. | |
4625 | * For instance, if the completion latencies are tight, we can | |
4626 | * get closer than just half the mean. This is especially | |
4627 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
4628 | * than ~10 usec. We do use the stats for the relevant IO size |
4629 | * if available which does lead to better estimates. | |
64f1c21e | 4630 | */ |
720b8ccc SB |
4631 | bucket = blk_mq_poll_stats_bkt(rq); |
4632 | if (bucket < 0) | |
4633 | return ret; | |
4634 | ||
4635 | if (q->poll_stat[bucket].nr_samples) | |
4636 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
4637 | |
4638 | return ret; | |
4639 | } | |
4640 | ||
c6699d6f | 4641 | static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc) |
06426adf | 4642 | { |
c6699d6f CH |
4643 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc); |
4644 | struct request *rq = blk_qc_to_rq(hctx, qc); | |
06426adf JA |
4645 | struct hrtimer_sleeper hs; |
4646 | enum hrtimer_mode mode; | |
64f1c21e | 4647 | unsigned int nsecs; |
06426adf JA |
4648 | ktime_t kt; |
4649 | ||
c6699d6f CH |
4650 | /* |
4651 | * If a request has completed on queue that uses an I/O scheduler, we | |
4652 | * won't get back a request from blk_qc_to_rq. | |
4653 | */ | |
4654 | if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT)) | |
64f1c21e JA |
4655 | return false; |
4656 | ||
4657 | /* | |
1052b8ac | 4658 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 4659 | * |
64f1c21e JA |
4660 | * 0: use half of prev avg |
4661 | * >0: use this specific value | |
4662 | */ | |
1052b8ac | 4663 | if (q->poll_nsec > 0) |
64f1c21e JA |
4664 | nsecs = q->poll_nsec; |
4665 | else | |
cae740a0 | 4666 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
4667 | |
4668 | if (!nsecs) | |
06426adf JA |
4669 | return false; |
4670 | ||
76a86f9d | 4671 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
4672 | |
4673 | /* | |
4674 | * This will be replaced with the stats tracking code, using | |
4675 | * 'avg_completion_time / 2' as the pre-sleep target. | |
4676 | */ | |
8b0e1953 | 4677 | kt = nsecs; |
06426adf JA |
4678 | |
4679 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 4680 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
4681 | hrtimer_set_expires(&hs.timer, kt); |
4682 | ||
06426adf | 4683 | do { |
5a61c363 | 4684 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
4685 | break; |
4686 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 4687 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
4688 | if (hs.task) |
4689 | io_schedule(); | |
4690 | hrtimer_cancel(&hs.timer); | |
4691 | mode = HRTIMER_MODE_ABS; | |
4692 | } while (hs.task && !signal_pending(current)); | |
4693 | ||
4694 | __set_current_state(TASK_RUNNING); | |
4695 | destroy_hrtimer_on_stack(&hs.timer); | |
1052b8ac | 4696 | |
06426adf | 4697 | /* |
c6699d6f CH |
4698 | * If we sleep, have the caller restart the poll loop to reset the |
4699 | * state. Like for the other success return cases, the caller is | |
4700 | * responsible for checking if the IO completed. If the IO isn't | |
4701 | * complete, we'll get called again and will go straight to the busy | |
4702 | * poll loop. | |
06426adf | 4703 | */ |
06426adf JA |
4704 | return true; |
4705 | } | |
06426adf | 4706 | |
c6699d6f | 4707 | static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie, |
5a72e899 | 4708 | struct io_comp_batch *iob, unsigned int flags) |
bbd7bb70 | 4709 | { |
c6699d6f CH |
4710 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie); |
4711 | long state = get_current_state(); | |
4712 | int ret; | |
bbd7bb70 | 4713 | |
aa61bec3 | 4714 | do { |
5a72e899 | 4715 | ret = q->mq_ops->poll(hctx, iob); |
bbd7bb70 | 4716 | if (ret > 0) { |
849a3700 | 4717 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4718 | return ret; |
bbd7bb70 JA |
4719 | } |
4720 | ||
4721 | if (signal_pending_state(state, current)) | |
849a3700 | 4722 | __set_current_state(TASK_RUNNING); |
b03fbd4f | 4723 | if (task_is_running(current)) |
85f4d4b6 | 4724 | return 1; |
c6699d6f | 4725 | |
ef99b2d3 | 4726 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
bbd7bb70 JA |
4727 | break; |
4728 | cpu_relax(); | |
aa61bec3 | 4729 | } while (!need_resched()); |
bbd7bb70 | 4730 | |
67b4110f | 4731 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4732 | return 0; |
bbd7bb70 | 4733 | } |
1052b8ac | 4734 | |
5a72e899 JA |
4735 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob, |
4736 | unsigned int flags) | |
1052b8ac | 4737 | { |
d729cf9a CH |
4738 | if (!(flags & BLK_POLL_NOSLEEP) && |
4739 | q->poll_nsec != BLK_MQ_POLL_CLASSIC) { | |
c6699d6f | 4740 | if (blk_mq_poll_hybrid(q, cookie)) |
85f4d4b6 | 4741 | return 1; |
c6699d6f | 4742 | } |
5a72e899 | 4743 | return blk_mq_poll_classic(q, cookie, iob, flags); |
bbd7bb70 JA |
4744 | } |
4745 | ||
9cf2bab6 JA |
4746 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4747 | { | |
4748 | return rq->mq_ctx->cpu; | |
4749 | } | |
4750 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4751 | ||
2a19b28f ML |
4752 | void blk_mq_cancel_work_sync(struct request_queue *q) |
4753 | { | |
4754 | if (queue_is_mq(q)) { | |
4755 | struct blk_mq_hw_ctx *hctx; | |
4756 | int i; | |
4757 | ||
4758 | cancel_delayed_work_sync(&q->requeue_work); | |
4759 | ||
4760 | queue_for_each_hw_ctx(q, hctx, i) | |
4761 | cancel_delayed_work_sync(&hctx->run_work); | |
4762 | } | |
4763 | } | |
4764 | ||
320ae51f JA |
4765 | static int __init blk_mq_init(void) |
4766 | { | |
c3077b5d CH |
4767 | int i; |
4768 | ||
4769 | for_each_possible_cpu(i) | |
f9ab4918 | 4770 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
c3077b5d CH |
4771 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4772 | ||
4773 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4774 | "block/softirq:dead", NULL, | |
4775 | blk_softirq_cpu_dead); | |
9467f859 TG |
4776 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4777 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4778 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4779 | blk_mq_hctx_notify_online, | |
4780 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4781 | return 0; |
4782 | } | |
4783 | subsys_initcall(blk_mq_init); |